Limited maintenance has resumed. We support the local Agent Panel and your choice of provider, including offline models. Releases focus on verified critical fixes. Automatic issue reporting remains disabled. We will reassess support when Comfy’s official local panel becomes publicly available.
Limited maintenance update
The updated support policy and the original sunset announcement.
Model highlights
A chronological tour of the open image & video models you can run locally in ComfyUI — what each is best at, and how to set it up in one command with a comfyui-mcp pack and the Panel. Oldest first, ending on the current open-weight champ, Ideogram 4.WAN 2.2
The open-source video king — how the high/low-noise experts actually work,
I2V/T2V, GGUF tiers, and longer videos.
Qwen-Image & Qwen-Image-Edit
20B text-to-image plus instruction editing and manual-mask inpainting, with a
WAN 2.2 refine combo.
WAN Animate 2.2
Drive a character with a reference video — v2v character & motion replace from
a single image.
WAN Transparent Expressions
Alpha-channel character expression sprites (looping WebP/GIF) for SillyTavern
and VTuber overlays.
Z-Image (Turbo & Base)
6B, runs under 8 GB — the low-VRAM speed king. Turbo vs Base, ControlNet, and
your own LoRAs.
ERNIE-Image
Apache-2.0 text-to-image with razor-sharp multilingual text rendering, under
8 GB VRAM.
ANIMA 1.0
A ~2B anime model that generates and trains under 6 GB — Danbooru tags plus
natural language.
LTX-2.3
Fast GGUF video (T2V/I2V) and the LTX Director timeline editor inside ComfyUI.
Krea 2 Turbo
Krea.ai’s 12B open-weight text-to-image — 8-step turbo speed, Qwen3-VL
prompting, and JSON area prompts.
Ideogram 4 — the finale
The 9.3B open-weight king of readable in-image text and design layout, with
JSON area prompting.
Video post-production
Upscaling video in ComfyUI
SeedVR2 & FlashVSR temporal restore, the downscale-first trick, the quick
per-frame ESRGAN path, and RIFE/FILM interpolation (built into ComfyUI 0.26).
Extending video with Pusa
Pusa 2.2 temporal flowmatching on the WAN 2.2 stack — continue a clip instead
of regenerating it, chain extensions, all on your own GPU.
Posts
comfy-kitchen in ComfyUI: see the faster path, then prove it
kitchen and panel_kitchen report what this GPU can run, name the UNETLoader
still on default dtype, and apply fp8 / INT8 attention / NVFP4 — or revert.
August 21, 2026
ComfyUI on RunPod: rent a cloud GPU
A first-party RunPod connector: one-tap deploy of a pod that runs YOUR exact
custom nodes and models, an honest local⇄pod host pill, live GPU/cost status,
and idle auto-stop — driven from the panel and the phone, no console or CLI.
July 21, 2026
Train a FLUX LoRA in the cloud on RunPod
No 24 GB card? Rent one. The LoRA trainer’s new SSH transport stands ai-toolkit
up on a RunPod pod and runs the same train_* flow on rented hardware — dataset
up over ssh, .safetensors back in models/loras/. July 21, 2026
Train a character LoRA by asking
The panel agent drives ostris’s ai-toolkit end to end through train_* tools —
preflight, dataset staging, streamed progress, honest cancel — verified with a
real FLUX.1-dev character run on a 4090. July 20, 2026
Blind mode: a privacy promise the agent can't break
A user proved our “the agent never sees the image” toggle was leaky. The fix:
enforcement at the tool-server boundary — one wrapper that withholds pixels
from every image-returning tool, current and future, mechanically rather than
promptably. July 20, 2026
GPT-5.6 Sol, Terra, and Luna land in the ComfyUI panel
The new frontier family on your ChatGPT subscription — max and ultra efforts
with per-model ceilings from the live catalog, account-aware deprecation that
never force-switches a runnable pin, and the SDK-pin war story behind tracking
a fast-moving provider. July 20, 2026
Kimi K3: the biggest open-weight model ever now drives your ComfyUI canvas
Moonshot’s 2.8T-parameter K3 — third on the intelligence boards, first on
Frontend Code Arena, weights landing July 27 — became a first-class provider
two days after release. Why it’s a big deal, the honest caveats
(always-max thinking, $15/M output), and the one-key setup. July 18, 2026
The 405 that wasn't: three dialects of ComfyUI-Manager
A debugging war story in three acts: per-operation routes vs task envelopes vs
a legacy UI hiding under /v2 — and the frontend catchall that turns every wrong
guess into a 405. July 18, 2026
Your conversation now survives workflow switching
The conversation is the unit of continuity; the workflow is just the canvas
target. How the orchestrator rebinds a live agent across workflow tabs — no
respawn, so even in-memory local models keep their history. July 17, 2026
CivitAI inside ComfyUI: browse, like, and run without a download folder
A full CivitAI browser in the sidebar — lightbox with generation details,
like/unlike with a synced likes collection, @creator search qualifiers, and
workflows that load straight onto the live canvas. July 17, 2026
The render finished while you were making coffee: ComfyUI on your phone
The mobile companion (iOS + Android beta): same agent, same providers, CivitAI
parity, one multi-purpose QR to pair — and tab-mirror remote control that lets
the phone drive a desktop session. July 16, 2026
Flatten the spaghetti: making expert workflows readable
Get/Set buses, Reroutes, and use-everywhere broadcasts resolved to direct
links IN PLACE — layout, groups, and widgets preserved — plus the
positional-widget bug the work exposed. July 15, 2026
The orchestrator that never goes stale
Hourly npm self-update and dev-build self-restart, coalesced to fire only when
every agent is idle and nothing is rendering — infrastructure that updates like
a browser, not a printer driver. July 14, 2026
We fine-tuned Gemma 4 to drive ComfyUI — and built an arena to prove it
A gemma4 ladder (e2b/e4b/12b) trained on server-verified trajectories, scored
by a best-of-3 arena of real ComfyUI tasks — including the regression we
published against our own model, the train/serve format mismatch behind it, and
the temperature-0 loop bug the arena couldn’t see. July 14, 2026
Giving every backend eyes — and honesty when a model doesn't have them
Vision is a property of the model, not the provider. Always attempt image
delivery, strip-and-retry once on rejection, and leave an in-history note so
the model can never pretend it saw an attachment it didn’t. July 14, 2026
Bring any LLM: free local models now drive ComfyUI
Ollama, LM Studio, llama.cpp, or any OpenAI-compatible endpoint — no account,
no API key, fully offline. The compact tool router that makes the full tool
catalog fit a
4B model, the loop-breakers small models need, and the VRAM pause that lets one
GPU run both the diffusion model and the LLM driving it. July 14, 2026
Two brains, one canvas: ComfyUI on Claude or ChatGPT
The ComfyUI sidebar agent now runs on EITHER your Claude or your ChatGPT
subscription — pick a provider, no API keys, full canvas-driving parity. How we
got there: an AgentBackend port, Codex over its app-server, panel tools as a
loopback MCP, and the schema gotcha that kept eating a tool. June 25, 2026
The self-healing ComfyUI agent: it fixes the node that crashed it
A Wan2.2 render hard-crashes ComfyUI at 99% with a native access violation. The
agent reads the crash dump on reload, names the culprit node, and escalates —
update → git pull → patch → offer a PR — plus install→restart→continue and
video-vision so it can judge its own output. June 25, 2026
The panel that talks back: an autonomous agent on your Claude subscription
Three tries to get the ComfyUI panel driven by a real background agent: channel
pushes can’t wake an idle session, the
--sdk-url transport that did everything
got locked down, and the Claude Agent SDK’s streaming input — persistent,
injectable, interruptible, subscription-authenticated, no API key — turned out
to be the gold standard. June 16, 2026Installer packs that can't rot: one manifest, every install
The community “1-click installer” .bat is great UX and a maintenance trap —
Windows-only, duplicated in a separate .sh, full of model URLs that rot
silently. How a single manifest drives both the double-click scripts and an
MCP-native install, with CI that fails the build the moment a model link dies.
June 16, 2026
From B to A: sharpening 70+ tool descriptions with TDQS
A case study using Glama’s Tool Definition Quality Score to find and fix the one weak tool that
was capping our whole server — plus five rules any MCP author can steal. May 25, 2026