In short
ComfyUI shipped 0.44.0 on November 4 and 0.45.0 on November 10, with two patches to 0.45.2. The open-model news is all MiniMax H3: P-H3-Image and H3-Image run natively, and within a day of H3-Image a pDMD LoRA cut its 6 evaluations to 2. HyperFlow H3 Flash brings a 14B video H3 to one 24 GB card. One default changed quietly: 0.45 turns on browser sign-in and partner-node tokens. Almost every H3 file inherits MiniMax's license, which excludes the EU, UK, Korea and the US.
If you post the same face every day, the last two weeks of ComfyUI come down to one model family. MiniMax H3 now runs natively as video, as P-H3-Image and as H3-Image, and the community turned each into something faster within days. The specific fact worth knowing first: H3-Image's stock checkpoint is locked to 6 sampling steps, and by November 10 a LoRA made it finish in 2. The October 31 briefing stopped at 0.43.1. This picks up from there.
Two weeks at a glance
Nov 2
LightX2V publishes Qwen-Image 2.1 Lightning v2 RL, a 4-step replacement DiT. License: Qwen Research, so not for a monetized feed.
Nov 4
ComfyUI 0.44.0: native P-H3-Image, an H3 latent upscaler, Sol-H3-VAE, H3 context windows, a VAE compute dtype node and ProRes saving. HyperFlow H3 Flash (14B) appears the same day.
Nov 5
PR #17540 merges: browser sign-in and partner-node tokens on by default, after 0.44.0 had reverted the first attempt.
Nov 6
H3 Dials: 14 KB steering files for MiniMax H3 and Pika H3 Flex, made from prompt pairs with no training.
Nov 9
NVIDIA's Open-H3-Regenerate-2K-IC: an in-context LoRA that re-renders a 768P H3 draft at 1080P, 2K or 4K.
Nov 10
ComfyUI 0.45.0 (42 changes): MiniMax-H3-Image, WeVisOne-VL, RelayVSR. 0.45.1 the same day adds the H3 Image partner node and LTX VFX nodes. The pDMD 2-step LoRA for H3-Image is published.
Nov 12
ComfyUI 0.45.2: a fix so some workflows load again. The changelog dates it Nov 12; the GitHub tag reads Nov 11.
Open models: what runs on your own card
0.44.0 made the H3 video stack lighter to run: a latent upscaler and Sol-H3-VAE, and context windows for H3, which let long clips render in overlapping pieces. It also brought P-H3-Image, Pruna's 8-step distilled image model, as a native model. We compared that one in MiniMax H3 Image vs P-H3-Image.
0.45.0 added MiniMax-H3-Image itself (#17655). It is a 33.1B model that samples in 6 steps without CFG and works best around 16 MP; our H3 Image review has the files and the 12 GB numbers a user posted. The same release added Tencent's WeVisOne-VL (text query in, boxes, masks and mattes out, Apache-2.0) and RelayVSR video super-resolution (Apache-2.0). For a character feed, WeVisOne-VL is the quiet win: a mask of "the jacket" or "the face" from one sentence, inside core, without a segmentation node pack.
| Change | Release | What it does | License |
|---|---|---|---|
| P-H3-Image | 0.44.0 | 8-step image model distilled from H3, native | MiniMax H3 community |
| H3 latent upscaler, Sol-H3-VAE | 0.44.0 | lighter H3 video upscale and decode | MiniMax H3 community |
| H3 context windows | 0.44.0 | long H3 clips in overlapping windows | core feature |
| VAE compute dtype node | 0.44.0 | pick the VAE precision per graph | core feature |
| MiniMax-H3-Image | 0.45.0 | 33.1B text-to-image and editing, 6 steps, no CFG | MiniMax H3 Image community |
| WeVisOne-VL | 0.45.0 | boxes, masks, mattes and OCR from a text query | Apache-2.0 |
| RelayVSR | 0.45.0 | video super-resolution | Apache-2.0 |
| Qwen 3.x | 0.45.0 | video input, combined linear weights for speed | Qwen license per model |
H3-Image in 2 steps: the pDMD LoRA
The stock H3-Image checkpoint is step-distilled to a fixed 6-step schedule, and ComfyUI's template sets it that way. On November 10 a pDMD 2-step LoRA appeared: rank 128 across attention and feed-forward in all 50 blocks, shipped as a ComfyUI-layout LoRA and as a 40 GB checkpoint with the LoRA already merged. Its card says both routes were verified on ComfyUI master of November 10, at about 20 seconds per 2048 by 2048 image on an H200.

Two cautions before you swap it into a daily graph. The LoRA repo is tagged Apache-2.0, but it is an adapter on MiniMax-H3-Image, and the base license still governs what you render with it. And the speed claim is the author's, on a data-center card: nobody has posted a same-seed face comparison of 6 steps against 2. That is the test worth running, below.
Video: Flash, Dials and a 2K regenerate pass
HyperFlow H3 Flash from Video Rebirth is H3 cut to 13.7B (13.9B for the reference-to-video variant), sampling in 4, 6 or 8 steps without CFG. Its card lists 92 seconds for a 5.2-second clip on an RTX 5090 at 6 steps, and a Q4_K_M file of 9.1 GB that runs on 24 and 32 GB cards. It says 16 GB cards are not measured yet, which is honest and worth taking literally.
H3 Dials steer H3 with 14 KB files built from opposite prompt pairs. The motion dial runs from 0.05 times normal motion at -2 to about 2.7 times at +2, which for a posing character is a cinemagraph slider. The author flags that the sinister mood dial crushes speech, with a Whisper word error rate of 0.84 at +2 against 0.00 at base, so keep mood dials off talking clips. Open-H3-Regenerate-2K-IC re-renders a 768P draft at up to 4K while keeping faces and lip sync, but its example command runs on 8 GPUs; treat it as a rented-machine finishing pass.
H3 derivatives this fortnight, in one card
- pDMD 2-step LoRA
- H3-Image in 2 evaluations; ~20 s at 2048x2048 on an H200 (author)
- HyperFlow H3 Flash
- 13.7B video + stereo audio; 92 s per 5.2 s clip on an RTX 5090; Q4_K_M 9.1 GB
- H3 Dials
- 14 KB, no training; motion 0.05x to 2.7x; mood dials can break speech
- Open-H3-Regenerate-2K-IC
- 768P draft to 1080P, 2K or 4K; example uses 8 GPUs
- Viggle Remake
- swap and edit LoRAs on H3; 12 to 24 GB by hand with int8 or GGUF
- Shared catch
- all carry the MiniMax H3 community license: EU, UK, Korea, US excluded without authorization
What creators are saying
One user hit an unknown node pack loading H3-Image and called the release rushed; a reply says the current stable ComfyUI runs it out of the box.
ComplaintComfy-Org/MiniMax-H3-Image discussion on Hugging Face
Editing with two reference photos of two people, one face comes out different from its source.
QuestionComfy-Org/MiniMax-H3-Image discussion on Hugging Face
For a detailer pass with the 2-step LoRA, a reply suggests LoRA strength 0.5 and 3 to 4 steps, which it calls just about passable.
A paired test found Portuguese speech articulation drops on HyperFlow Flash compared with full H3; another user says Spanish too. The authors call it a capacity trade-off and suggest 8 steps.
The default that changed: sign-in on
0.44.0 turned on local_web_sso and partner_node_token by default and then reverted it in the same release. PR #17540 brought it back on November 5, and it ships in 0.45.0. Per the PR, browser sign-in only activates on non-Cloud loopback hosts, and partner-node calls switch to minted tokens, falling back to your workspace token if minting fails. Nothing here sends your local renders anywhere, but if you run a locked-down box and want no sign-in prompt, the PR documents --feature-flag local_web_sso=false.
Partner nodes: billed per run
The paid list grew again: H3 Max editing (recast, extend, reframe, insert, lip sync, camera control), Pika H3 Flex, Kling 4.0 Flash, GPT Image 2.6 with GIF and SVG nodes, Seedream 5.5, Seedance 2.5 at 4K, Gemini Omni Flash 1.2, LTX VFX nodes, and an H3 Image partner node in 0.45.1. Seedance 1.x nodes are deprecated and Grok's image models retired. These spend Comfy Credits, not your GPU. Fine for one hero shot; a weak base for a recurring character, since you cannot pin a version or train on it.
What to do this week
- 1
Update to 0.45.2, not 0.45.0
0.45.2 fixes workflows that failed to load. Open the H3-Image template from the template list instead of rebuilding the graph; the missing nodes guide covers the unknown-node error from the thread above.
- 2
Run the 6 against 2 test on your own character
Same prompt, same seed, same references: stock H3-Image at 6 steps, then the pDMD LoRA at 2. Judge eyes, skin and hairline at 100% before you trust the speed.
- 3
Decide on the license before the file
If your account earns in the EU, UK, Korea or the US, every H3 file in this briefing needs MiniMax's authorization. WeVisOne-VL and RelayVSR are Apache-2.0 and safe to adopt now.
- 4
Keep mood dials off talking clips
Use the motion dial for posing loops; test any mood dial with speech before posting.
If you test these on your own install, Musetta can help: it runs on your own ComfyUI, installs the nodes and models a template is missing, and keeps each character's reference photos together, so a 6 against 2 step test is one request. The manual route is in our consistent character guide.

Questions
- What is the newest ComfyUI version?
- As of November 14, 2026 the newest release in the ComfyUI changelog is v0.45.2, dated November 12, which fixes some workflows that failed to load. v0.45.0 and v0.45.1 came out on November 10, and v0.44.0 on November 4.
- Does ComfyUI 0.45 support MiniMax H3 Image natively?
- Yes. v0.45.0 added MiniMax-H3-Image support (PR #17655), and Comfy-Org hosts repacked files. The model samples in 6 steps without CFG; a community pDMD LoRA brings it to 2 evaluations.
- How do I turn off the new sign-in in ComfyUI 0.45?
- 0.45.0 turns on local_web_sso and partner_node_token by default (PR #17540). The PR documents starting ComfyUI with --feature-flag local_web_sso=false to disable browser sign-in. It only activates on local loopback hosts.
- Can I use MiniMax H3 derivatives for a monetized account?
- Check where you operate. The MiniMax H3 and H3 Image community licenses exclude the EU, UK, Republic of Korea and the US without authorization, and LoRAs and distillations built on H3 inherit that base license whatever their own tag says.
Sources
- 01ComfyUI changelog · Comfy Org, 2026-11-12
- 02ComfyUI v0.45.0 release notes · Comfy Org on GitHub, 2026-11-10
- 03ComfyUI v0.44.0 release notes · Comfy Org on GitHub, 2026-11-04
- 04PR #17540: turn on local_web_sso and partner_node_token by default · Comfy Org on GitHub, 2026-11-05
- 05MiniMax-H3-Image pDMD 2-Step LoRA · Hugging Face, 2026-11-10
- 06MiniMax-H3-Image model card and license · MiniMax on Hugging Face, 2026-11-10
- 07HyperFlow H3 Flash (14B) · Video Rebirth on Hugging Face, 2026-11-04
- 08H3 Dials · Hugging Face, 2026-11-10
- 09Open-H3-Regenerate-2K-IC · NVIDIA on Hugging Face, 2026-11-09
- 10Viggle Remake · Viggle on Hugging Face, 2026-11-07
- 11ComfyUI news · ComfyUI Wiki, 2026-11-14
- 12Comfy-Org/MiniMax-H3-Image discussion 14 · Hugging Face community, 2026-11-11
- 13HyperFlow H3 Flash discussion 6 · Hugging Face community, 2026-11-12
Researched and written by the Musetta editorial desk with AI assistance, from the sources above. Numbers come from those sources; nothing is benchmarked by us unless the article says so. Spotted an error? Tell us on Discord and we fix it.