COMFYTAM · RESEARCH AND PRACTICAL GUIDES

The ComfyUI ecosystem, in context.

Explore the project from its first commit in January 2023 through September 2026: dated development records, attributed company announcements and practical workflow questions. The restored visual dashboard brings the original ecosystem layout together with the newer evidence-led research, without inventing a market model.

By Professor Dominic Marrocco | First published 1 July 2025 | Merged revision 1 October 2026 | Selected announcement cutoff: 30 September 2026

Compare installation routesExplore workflow resources
45 Calendar months, Jan 2023 to Sep 2026
15 Existing monthly chapters retained
4 Repositories in the dated snapshot

A repository snapshot you can inspect

Stars in a fixed four-repository sample, observed 16 September 2026. This is a dated snapshot, not a historical growth series, unique-user census, download total or market-share comparison.

Accessible underlying observations; no overlapping counts are summed
RepositoryStarsForksObservation (UTC)Evidence
Comfy-Org/ComfyUI13341315783GitHub API
Comfy-Org/ComfyUI_frontend2013699GitHub API
Comfy-Org/ComfyUI-Manager161512481GitHub API
Comfy-Org/workflow_templates904188GitHub API

Download the dated snapshot; read the snapshot methodology. The API shows current metadata when visited later; the retained JSON records the actual observation. Independent forks, node packs and private deployments are outside this fixed sample.

Financing disclosures, not a market-size chart

Two company announcement amounts are shown on the same USD-million scale. The $500M April valuation and $47M reported total are different quantities and are deliberately excluded from the bars.

Selected Comfy financing announcementsSeventeen million dollars announced 16 September 2025; thirty million dollars announced 24 April 2026. Disclosures are not an independently verified cap table.16 Sep 2025$17M24 Apr 2026$30M0102030Announced funding amount (USD millions)
Attribution and limits for the plotted amounts
DateAmountNamed backersPrimary source
16 September 2025$17M; valuation unstatedPace, Chemistry, Abstract; Essence VC acknowledgedComfy disclosure
24 April 2026$30M financing; $500M valuation; $47M total reportedCraft (lead), Pace, Chemistry, TruArrow and unnamed othersComfy disclosure; Craft, 25 April

The announcement dates are not established transaction closing dates. The valuation basis and allocations remain unknown. Secondary reporting describes a different earlier-round amount and date; this unresolved discrepancy remains in the full funding register, not an invented extra round. No investment recommendation is supplied.

From the first commit to September 2026

Comfy dates the first commit to 16 January 2023 in its second-anniversary retrospective. Earlier experiments in October 2022 are described in the September 2025 funding post; they are not a January 2023 company-incorporation date. Company retrospective, 16 January 2025; company origin account, 16 September 2025. Project origins, company formation and financing announcements are distinct events.

This calendar represents all 45 months from January 2023 through September 2026. The first 30 entries use up to five end-of-month core-repository commits each, retrieved on 1 October 2026. They are selected evidence, not exhaustive release histories or monthly activity counts. The later 15 retain their published methods chapters and company-source commentary. Their different evidence types are deliberately labelled.

Repository record / 2023-01

January 2023: Graph primitives

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Save a small graph using the latent transform or advanced sampler you actually need. Retain its input dimensions and seed; node availability alone does not establish interchangeable outputs.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2023-02

February 2023: Memory and backend handling

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Distinguish CUDA and ROCm environments before investigating a memory failure. Record the installed backend and a failing workload instead of interpreting cache cleanup as extra physical VRAM.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2023-03

March 2023: Model patching and output history

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Compare a patched model with an unchanged baseline. Keep output-history records together with the graph and model revision so a saved image can be traced to its settings.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2023-04

April 2023: Conditioning and checkpoint defaults

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Make checkpoint fallback and conditioning weights explicit in a shared graph. A successful load through a default checkpoint is not evidence that the intended checkpoint was selected.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2023-05

May 2023: Custom-node registration

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Record node-pack revisions before changing registration code. Check how a missing extension is reported and retain a recoverable environment; this is a proposed check, not a security certification.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2023-06

June 2023: Model merging and loader caching

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Distinguish a cached model from a newly loaded revision when comparing outputs. Preserve merge settings and source-model terms; a merge feature does not grant permission to redistribute weights.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2023-07

July 2023: Latent output and execution resources

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Keep latent outputs and ordinary image files separately labelled. A downstream collaborator needs the correct decoder and graph context; a file saved successfully is not automatically a portable result.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2023-08

August 2023: Tiled decoding and controls

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Record tile settings, decoder revision and output dimensions when diagnosing memory pressure. Compare the complete delivered image rather than treating a lower-memory decode as an equivalent quality result.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2023-09

September 2023: Sampler and scheduler choices

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Freeze the sampler, scheduler, seed and step count as one condition. A newly available scheduler is a distinct experimental choice, not evidence of universal improvement over existing settings.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2023-10

October 2023: Attention and image quantisation

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Separate attention-memory handling from image-processing changes. Retain unsuccessful runs as well as outputs so a patch intended to recover a workload is not silently presented as a speed benchmark.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2023-11

November 2023: Grouped graphs and model recipes

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Check that grouped nodes retain understandable inputs and outputs when reopened. Use a documented model-specific recipe instead of assuming an SDXL Turbo workflow accepts ordinary sampling settings unchanged.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2023-12

December 2023: Validation and CPU decoding

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Exercise invalid node inputs before distributing a graph. If decoding runs on CPU, record that execution path alongside elapsed time; do not compare it with a GPU-only result without qualification.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2024-01

January 2024: Interface and group management

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Test an edited mask and grouped graph after saving and reopening. Pin the framework environment separately from interface preferences, because a frontend improvement does not establish backend compatibility.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2024-02

February 2024: Model-specific sampling

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Document the model and sampling configuration together. Support for a model family or continuous sampling path does not prove that a previously saved graph has the correct conditioning and settings.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2024-03

March 2024: Platform defaults and import diagnostics

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Keep platform-specific precision and extension-import logs with the workflow record. A changed default on one operating system must not be advertised as a universal accelerator improvement.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2024-04

April 2024: Upscale sampling and mask handling

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Keep mask batches, upscale settings and scheduler configuration in the same test record. Validate delivered dimensions and alignment; a sampler feature is not a substitute for checking the final composition.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2024-05

May 2024: Preview and upscaler dependencies

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Separate a preview from the exported deliverable and record the upscaler dependency. Inspect deterministic settings for the actual backend before treating visually similar previews as reproducible output.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2024-06

June 2024: Audio workflow records

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Retain an audio file's embedded workflow and inspect the exported file as well as its preview. Confirm the model licence and output format independently; an import feature does not establish commercial rights.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2024-07

July 2024: Text-encoder precision and memory

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Pin text-encoder precision and model identity alongside the GPU environment. Compare a fixed prompt and model revision before attributing a changed result to memory management alone.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2024-08

August 2024: Frontend transition and Flux

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Company announcement, 2024-08-06: The team announced Flux support, stable releases and a TypeScript frontend transition planned for 15 August. Its stated priorities were development speed and extension compatibility. The announcement's planned execution-inversion work is not treated as already released. Read the dated primary source.

Operator check, proposed: Keep interface and core revisions separately identifiable during the transition. Reopen a small extension-dependent graph before updating a production collection, and retain a rollback route.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2024-09

September 2024: Compilation and low-memory adapters

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Compare eager and compiled execution under the same environment and workload. Record adapter format and memory mode; compile-path fixes are not evidence that every custom extension is compatible.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2024-10

October 2024: Desktop and node compatibility

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Company announcement, 2024-10-21: The V1 announcement described a packaged Desktop experience and registry-based extension management. The desktop invitation was a closed beta, not proof of general availability or an independently verified security guarantee. Read the dated primary source.

Operator check, proposed: Treat desktop packaging, accelerator settings and individual node behaviour as separate acceptance checks. Preserve an exported workflow so a convenient installation route does not become the only recovery path.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2024-11

November 2024: Inpainting and video previews

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Inspect the actual inpaint or video deliverable, not only its preview. Experimental accelerator instructions remain conditional; retain device and framework versions when deciding whether they apply.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2024-12

December 2024: Sampler errors and extension contracts

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Keep a minimal failing graph when a sampler or custom node fails. Distinguish error handling and type-hint changes from a feature launch; a clearer failure is useful without implying a successful run.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2025-01

January 2025: Frontend and hardware guidance

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Company announcement, 2025-01-16: The second-anniversary post set out an operating-system-for-generative-AI ambition, with interface, team workflow and dependency-management priorities. This is a company strategy statement, not evidence that those priorities had all shipped. Read the dated primary source.

Operator check, proposed: Check new hardware instructions against the installed framework build before upgrading. Preview and batching changes need a saved acceptance example rather than an assumed speed or capacity increase.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2025-02

February 2025: Wan integration and precision modes

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Pin model revision, permitted inputs and precision when trying a video graph. An integration commit establishes repository work, not measured video quality, broad device support or a complete release census.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2025-03

March 2025: Video preprocessing and latent channels

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Preserve preprocessing settings and expected latent shape with the workflow. Compare the final sequence, including frame count and dimensions, before interpreting a successful graph execution as an equivalent result.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2025-04

April 2025: Offloading and update tests

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Keep an untouched environment before testing updater or offloading changes. Record model-loading and generation observations separately; an automated-test change does not certify every user's installation.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2025-05

May 2025: Templates and API paths

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Retain the template version and inspect API route handling in the actual deployment. Check a known input and failure response rather than assuming a template update is a backwards-compatible production interface.

Retrospective written 1 October 2026. Source-calendar entry.

Repository record / 2025-06

June 2025: Sampling optimisation and release process

Selected dated commits in the core repository document the following work. These are change records, not a claim that every change shipped in a stable release, passed a benchmark or reached all users.

Operator check, proposed: Compare an unchanged graph before and after an optimisation. Release-workflow edits and dependency updates are repository changes, not evidence of product adoption, revenue or a completed security review.

Retrospective written 1 October 2026. Source-calendar entry.

July 2025 onwards: fifteen operator decisions

These notes are a practical companion to the longer source calendar, not a second copy of it. Each month asks what an installer, workflow author or team should check before relying on an announcement. The suggested checks below have not been executed by this review. Dates belong to the cited announcements; the initial commentary was written in September 2026; calendar alignment and the September close-out were reviewed on 1 October 2026.

July 2025: establish a recoverable baseline

Comfy's 28 July Wan2.2 announcement provides a dated starting point for a video-workflow record. It does not validate the historical chapter's commercial estimates. Keep the original chapter as a corrected historical document, not the denominator for a growth calculation.

Operator check: before changing a working installation, retain one permitted input, its graph, the required model references and the output you expect to reproduce. Record what cannot be redistributed so a collaborator does not mistake a workflow file for a complete portable package.

Comfy announcement, 28 Jul 2025 · Historical chapter and correction

August 2025: define the subgraph contract

The 7 August subgraph release makes reusable graph components a concrete subject for review. The post distinguishes a latest-commit release from a forthcoming stable version; a recipient's installed build still matters.

Operator check: write down a component's inputs, outputs and dependencies before sharing it. Reopen it in the recipient's environment and deliberately leave out a dependency to examine the failure message. The useful question is whether the component is understandable and recoverable, not whether a tidier canvas proves faster execution.

Comfy announcement, 7 Aug 2025 · August source chapter

September 2025: translate funding into questions

Comfy disclosed $17M on 16 September, naming Pace Capital, Chemistry and Abstract Ventures and acknowledging Essence VC. Its priorities included local usability, Cloud and ecosystem stability. No valuation or precise transaction closing date is established by that post.

Operator check: maintain separate columns for promised work, released functionality and your own acceptance evidence. A financing announcement is not a service guarantee. Before committing a project, identify an export route and the person responsible for recovering the workflow if a dependency or hosted service changes.

Company disclosure, 16 Sep 2025 · September source chapter

October 2025: mark the network boundary

Sora 2 Partner Nodes were announced on 8 October. A node appearing inside a familiar graph editor does not mean its model runs on the local machine, nor does it establish permission to send a client's material to a hosted service.

Operator check: label the stages that leave the device, what they transmit and who authorises that transmission. Use permitted test material and retain the provider's job identifier. Separate remote charges and failures from local GPU observations so a mixed graph does not produce a misleading local-performance comparison.

Comfy announcement, 8 Oct 2025 · October source chapter

November 2025: rehearse the cloud handoff

Cloud entered public beta on 4 November. The announcement also described future APIs, parallel execution and collaboration. Keep those plans separate from what was available in that beta; later releases do not retroactively establish November capability.

Operator check: move a small, known workflow rather than an entire production library. Compare model availability, allowed nodes, output retrieval and failure recovery. Keep your local version intact until the exported output and dependency record have been reviewed. Reduced installation effort is not evidence that every custom workflow transfers unchanged.

Comfy announcement, 4 Nov 2025 · November source chapter

December 2025: verify upstream identity

The 29 December post proposed moving the main repository into Comfy Org by 6 January. Treat it as an announcement of intent, not proof that a migration happened on the proposed date or that the software received a security certification.

Operator check: inspect the official repository destination, release reference and dependency lock before updating automation. Document any changed source URL alongside the revision it resolves to. Do not replace a pinned dependency merely because a new organisation name looks more official; provenance and executable changes need separate examination.

Comfy announcement, 29 Dec 2025 · December source chapter

January 2026: keep hardware claims conditional

The 6 January AMD announcement and 9 January NVIDIA optimisation report describe different hardware and software prerequisites. They should not be combined into a universal speed comparison or a promise that every supported installation can run the same workload.

Operator check: record the GPU, driver, runtime, framework build and model precision as one configuration. Compare one unchanged workload before and after an update, retaining unsuccessful runs. Label upstream benchmark observations separately from your results; an optimisation intended for one accelerator can be irrelevant or harmful on another configuration.

AMD announcement, 6 Jan · NVIDIA report, 9 Jan · January source chapter

February 2026: measure rather than fill the silence

No February corporate strategy or financing announcement is established in this selected source register. That is an evidence gap, not a finding that the company was inactive. March announcements remain in March rather than being moved here to make the calendar appear busier.

Operator check: use the retained memory-pressure method to distinguish model loading, sampling and output decoding. Save peak-memory observations and the settings that produced them. A failure log can answer a hardware-sizing question more honestly than an invented monthly adoption statistic. This is a proposed measurement task, not a completed February benchmark.

February method and evidence gap

March 2026: test the controls you hide

Comfy announced Cloud's exit from beta on 4 March and App Mode, App Builder and a Hub preview on 10 March. A smaller interface creates a different user experience, but it does not remove the underlying graph's assumptions.

Operator check: list which parameters the app exposes and which remain fixed. Try invalid inputs, an interrupted job and a second user with different permissions. Confirm who receives the output and who can recover the original graph. A successful demonstration for its author is not yet acceptance evidence for another operator.

Cloud announcement, 4 Mar · App announcement, 10 Mar · March source chapter

April 2026: track delivery, not a valuation proxy

On 24 April Comfy reported $30M financing, a $500M valuation and $47M total funding. Craft led, with Pace Capital, Chemistry, TruArrow and other participants. Craft published its investment rationale the next day. These are attributed disclosures, not a 42.uk valuation or a measure of market size.

Operator check: turn the stated local, cloud, collaboration and stability priorities into a dated delivery checklist. Neither valuation basis nor investor allocations should be inferred. The full register retains the unresolved earlier-round discrepancy instead of inventing an extra round or silently adjusting the disclosed total.

Company disclosure, 24 Apr · Investor statement, 25 Apr · April source chapter

May 2026: separate integrations by task

The 14 May model announcement and 27 May Krea Partner Node announcement cover different tasks and execution routes. A longer model list is not a common quality score: segmentation, video editing and hosted image creation need different acceptance criteria.

Operator check: give each integration a small input set appropriate to its actual purpose. Record where execution occurs, which settings are available and how costs are incurred. For a hosted route, preserve provider and model identifiers; for a local route, preserve the model revision and dependencies. Do not rank unlike tasks using one appealing example.

Model announcement, 14 May · Partner Node announcement, 27 May · May source chapter

June 2026: make agent actions reviewable

The 22 June mobile example was presented as a lightly tested prototype. The 29 June MCP announcement was a public beta. These are evidence of new interfaces to workflows, not proof of production hardening across every client or permission boundary.

Operator check: begin with read-only discovery, then authorise a bounded test job using non-sensitive inputs. Record the submitted graph, requested operation and resulting job identifier. Check cancellation, failed-job handling and access revocation before offering unattended execution. This proposed exercise makes the approval boundary visible without treating a beta label as a safety guarantee.

Prototype account, 22 Jun · MCP announcement, 29 Jun · June source chapter

July 2026: plan for the departing collaborator

The UAL partnership announcement on 16 July and Team plan announcement on 23 July make institutional workflows a useful review topic. Some broader education and team controls were described as forthcoming; they should not be recorded as already delivered.

Operator check: assign ownership for workflows, billing and retained assets before adding members. Rehearse a project ending or a collaborator leaving: identify what access must be removed and which materials the remaining team still needs. The presence of shared credits alone does not demonstrate granular budget control or a complete offboarding process.

Partnership announcement, 16 Jul · Team announcement, 23 Jul · July source chapter

August 2026: keep permissions and licences separate

Local MCP was announced on 18 August, followed by a MiniMax H3 commercial-licence reseller announcement on 27 August. Local execution, permission to install software and rights to use a model commercially are three different questions.

Operator check: distinguish inspecting installed resources from changing them or choosing a paid remote service. Before a commercial job, review the exact licence and intended use; the announcement describes exclusions rather than unrestricted resale rights. Retain the approved execution route with the workflow so a later agent cannot silently turn a local task into a different commercial arrangement.

Local MCP announcement, 18 Aug · Licence announcement, 27 Aug · August source chapter

September 2026: from provider choice to deployment

Forward Deployed Creatives on 4 September concerns team handover; Router on 23 September concerns explicit provider selection. Comfy's 30 September API announcement then describes versioned workflow Builds and managed deployment. These are distinct announcements, not a claim that hosted routes are interchangeable.

Operator check, proposed: retain the graph and its dependency revisions, distinguish a Build from its immutable release, and examine a recoverable failure before exposing an endpoint. Verify usage charges, permissions and portable recovery for the actual plan; the announcement is not an independently executed acceptance test.

4 September service announcement | 23 September Router announcement | 30 September API announcement | September source chapter

Three layers to compare

This is an editorial framework for organizing decisions, not a market-share model.

1. The workflow engine

Start with the workflow: models, nodes, inputs and outputs. Keep the graph and its dependencies alongside each result so another person can reconstruct the run.

2. Installation and management

Compare how each route installs, updates and isolates dependencies. An easier installer does not establish model compatibility or faster generation.

3. Products and services

Assess the actual task a product solves, where it runs, and what it costs. Require the same saved workflow and hardware conditions before comparing output or speed.

Installation: start with current primary sources

Use the 42.uk installation matrix to compare documented routes and their evidence dates. A listed platform is not proof that every model or custom node works there.

A result someone else can reproduce

  1. Save the exact workflow. Include model names and hashes, custom-node revisions, input files you can share, and the seed.
  2. Record the environment. Note the GPU and VRAM, operating system, driver, Python, framework and ComfyUI versions.
  3. Hold the workload constant. Keep resolution, batch size, steps, sampler and precision fixed. Report cold model loading separately from warmed generation.
  4. Retain every run. Report failures and memory errors alongside elapsed time and peak memory. A screenshot helps explain the setup; raw rows let readers check the arithmetic.
  5. Describe the limit. A successful run establishes that configuration only. A planned test, estimate or vendor claim must stay labelled as such.

Open the low-VRAM testing guide and recording template. The blank template is a way to collect evidence, not a dataset of measured results.

Commercial opportunity is a research question

The 2025 thesis was that installation, workflow packaging and managed execution might reduce practical friction. That remains an editorial hypothesis. This page does not establish addressable-market size, revenue potential, conversion rates or product leadership.

To test that hypothesis, define a user task, measure time to a successful first result, record support effort, and compare actual willingness to pay. Separate a download, a GitHub star, an active user and a paying customer; none is an interchangeable measure.

Dependencies and rights need their own review

Record the source and version of each installed component. Review the actual terms attached to the application, model, node and input material for your intended use. This guide is not a security audit, licence clearance or compliance certification.

A package manager can simplify setup; its presence alone does not demonstrate that third-party code has been audited. Review the relevant repository and its changes before adding a dependency.

Useful evidence to publish next

A small, reproducible comparison is more useful than an unsupported league table: one task, two documented installation routes, the same workflow, and a transparent log of successful and failed runs. Publish the configuration and observations before drawing a general conclusion.

Monthly research through September 2026

Archive expanded 29 September 2026. Read these as dated research and commentary, not a comparable adoption time series. September includes the 30 September API announcement; inclusion does not mean every product claim has been independently tested.

All months, funding and public identity sources

Common questions

Which installation route should I use?

Start with your operating system and GPU, then match them to current official documentation. Use a small known workflow to verify your particular setup before importing a large collection of custom nodes.

Does this page prove one product is faster?

No. This revision removes comparisons that lacked reproducible run evidence. Use the recording template to collect comparable measurements for your own hardware and workload.

Is the GitHub page the complete application source?

The linked GitHub repository contains the library information page. The research article is hosted on 42.uk; its GitHub information page states that scope explicitly.