Diffusion Desk

Open-source image & video AI, sourced daily

Diffusion Desk

News about ComfyUI, Civitai, civitai.red and new image & video models

Last updated 46 stories 33 sources

Image: civitai.com
Top storyModels

Muse by Stable Yogi v4.0 Extended: photoreal Krea 2 checkpoint

The most downloaded checkpoint of the month in Civitai's public API (SFW filter) got a v4.0 “Extended” update: more real skin, hair and fabric texture and more natural poses. Versions come in Q8 GGUF (Oct 21), FP8 and INT8 (Oct 23); quick settings are 8 steps, CFG 1, Euler.

CheckpointBase: Krea 2Licence: not stated

Note: Ranking from civitai.com/api/v1/models (Most Downloaded, month, nsfw=false) on Oct 26. The model page itself dates from June; v4.0 is the new release.

Source: Civitai model page

Image: huggingface.co
Models

MiniMax H3 Singularity v2: pruned int8, w6a8 and w4a8 files for ComfyUI

WAS repackaged the ref2va v2 release of the community Singularity fine-tune of MiniMax H3 as pruned, quantized files (int8 21.0 GB, w6a8 16.0 GB, w4a8 11.8 GB) in the layout of ComfyUI's own pruned H3 releases. Load them with Load Diffusion Model. The card publishes PSNR comparisons against the bf16 release.

Video checkpointBase: MiniMax-H3 (Singularity fine-tune)Licence: MiniMax H3 Community License

Source: Hugging Face model card

Image: huggingface.co
Models

MiniMax-H3 PTD LoRAs: 8- and 4-step reference-to-video

Taobao's content AI team distilled MiniMax-H3's Ref2VA task (reference image to video and audio) with Parametric Trajectory Distillation, releasing 8-step and 4-step accelerator LoRAs that use one adapter per step. A sibling repo, PTD-RefAsymOPD (Oct 22), adds 8- and 6-step versions with better sharpness and realism.

Video LoRABase: MiniMax-H3Licence: MiniMax H3 Community License

Note: Quality claims for RefAsymOPD come from the authors.

Source: Hugging Face model card

Image: huggingface.co
Models

Kromoss: a full MOSS-Image fine-tune with 8-step turbo LoRAs

The lodestones account published Kromoss, a full 1024px fine-tune of MOSS-Image-Base trained on 704k images. It is a drop-in replacement for MOSS-Image-SFT and ships two 8-step, CFG 1 turbo LoRAs distilled with DMD and DMD2.

CheckpointBase: MOSS-Image-BaseLicence: Apache 2.0

Note: Released two days after the MOSS-Image weights; quality comparisons on the card are the author's own.

Source: Hugging Face model card

Image: github.com
ComfyUI

ComfyUI v0.42.0 adds native MOSS-Image support

MOSS-Image now runs natively, with text-to-image and layout prompts set through bounding boxes. The release also speeds up Qwen 3.x, adds Qwen3.5 embeddings and ARM Linux tests, and bumps the frontend to 1.55.18. v0.42.1 (Oct 19) fixed crashes in the MOSS templates.

Source: ComfyUI GitHub releases

Image: huggingface.co
Models

Bria Fibo 2.0 Dev: one open model for generation and editing

Bria's Fibo 2.0 Dev is the full, non-distilled base of the new FIBO generation, unifying text-to-image and instruction-based editing (masked and multi-reference) in one model trained only on licensed data. It keeps structured VGL captions and claims better text rendering, texture and anatomy; a few-step Edit Turbo variant also exists.

Base modelBase: FIBO 2.0 (new base)Licence: CC BY-NC 4.0 (gated)

Note: Licence read from the repo metadata (non-commercial). The “#1 in editing among open-weights models” ranking is Bria's own claim, citing Artificial Analysis. ComfyUI v0.42.3 added Bria FIBO 2.0 partner nodes.

Source: Hugging Face model card

Image: civitai.com
Civitai

MOSS Image arrives on Civitai: long prompts and JSON layout

OpenMOSS's 5B open-weights text-to-image model is in the generator in SFT and Base versions. It takes prompts of up to 4K tokens, supports JSON layout and outputs 512px to 2K. Weights are Apache 2.0, and LoRA training is available in Training Studio.

Source: Civitai article

Image: huggingface.co
Models

FastH3-QAD-v2: 8-step MiniMax H3 video in NVFP4 on one GPU

FastVideo released an 8-step video-and-audio model with NVFP4 weights trained by quantization-aware distillation against a BF16 teacher. It targets a single Blackwell GPU (RTX 5090, RTX PRO 6000) or a DGX Spark and uses block-sparse VSA attention.

Video checkpointBase: FastH3 8-step v2 (MiniMax H3)Licence: not stated

Note: The card does not name the licence (“other”); the MiniMax H3 base is our reading of the FastH3 name and card.

Source: Hugging Face model card

Models

Lenovo UltraReal LoRA gets a MOSS Image version

Danrisi's photoreal style LoRA, the most downloaded LoRA of the month in Civitai's public API (SFW filter), added a v1.0 build for MOSS Image on Oct 17, after Qwen 2.1 (Sep 22) and Krea 2 versions. The same author also posted MOSS versions of NiceGirls UltraReal and 2000s Analog Core.

LoRABase: MOSS ImageLicence: not stated

Note: No image: the MOSS version's previews are not rated PG on Civitai. Ranking from the public API on Oct 26.

Source: Civitai model page

Image: civitai.com
Civitai

Flux.3 Klein on Civitai: edit with up to 10 references

BFL's 5.6B open-weights Flux.3 Klein (distilled and Base) is now in the generator for text-to-image and editing with up to 10 reference images. The int8 files run on an 8 GB GPU. The API-only FLUX 3 Image has been available since Oct 2.

Source: Civitai article

Image: civitai.com
Models

Krea 2 Identity Edit V2: one photo in, that person anywhere out

This Civitai LoRA for Krea 2 now ships as two LoRAs: Identity Edit V2 puts the person from one reference photo into any scene, pose or outfit with better likeness, and General Edit V2 handles plain-language edits such as removing objects, relighting, recolouring or swapping backgrounds.

LoRABase: Krea 2Licence: not stated

Note: A solo hobby project. Civitai lists no licence name, only its usage permissions.

Source: Civitai model page

Image: blog.comfy.org
ComfyUI

FLUX 3 Image Klein runs natively in ComfyUI with 5.6B open weights

Black Forest Labs' compact open-weights FLUX 3 model ships in base and distilled versions. It supports bounding-box layout, multi-reference editing and text rendering. Comfy reports 4.3x to 9.3x faster 1024px generation using INT8 quantization with the step-distilled model.

Source: Comfy blog

Image: blog.comfy.org
ComfyUI

Comfy Org launches a $25,000 open-source build challenge

Higgsfield promised to open-source its platform, so Comfy is inviting builders to make open versions of tools like Genjutsu and Cinema Studio on the Comfy Developer Platform. The challenge runs for two weeks, with a $10k grand prize and three RTX 5090s. Runpod, LTX, BFL and MiniMax are backing it.

Source: Comfy blog

Image: huggingface.co
Models

FLUX 3 Image Klein weights: 5.6B, distilled and Base checkpoints

Black Forest Labs published FLUX 3 Image Klein on Hugging Face: one 5.6B model for text-to-image, single-reference editing and multi-reference composition with up to 10 images, using Qwen3-VL-4B-Instruct as text encoder. The distilled checkpoint runs in 16 steps without guidance; Klein Base (50 steps, guidance 3.5) is meant for fine-tuning and LoRA training. BF16, FP8 and NVFP4 files are included.

Base modelBase: FLUX 3 (new base)Licence: FLUX Kommunity License (gated)

Note: The card says a prompt upsampler is required: short prompts give generic results. Date = Hugging Face repo creation.

Source: Hugging Face model card

Image: huggingface.co
Models

MOSS-Image Base and SFT: Apache 2.0 text-to-image weights

OpenMOSS released two open-weights text-to-image checkpoints, MOSS-Image-Base and the fine-tuned MOSS-Image-SFT, under Apache 2.0, with Diffusers support and an online demo. It is one of the most trending text-to-image repos on Hugging Face, and community fine-tunes and LoRAs are already appearing.

Base modelBase: MOSS-Image (new base)Licence: Apache 2.0

Source: Hugging Face model card

Models

DaSiWa MiniMax H3 Hybrid v4: a community video checkpoint

Darksidewalker's fine-tuned MiniMax H3 checkpoint reached Hybrid v4, compatible with both reference-to-video (REF2VA) and first/last-frame (FL2VA) workflows, with a 4-step Turbo variant. It is the top MiniMax H3 checkpoint in Civitai's monthly download ranking.

Video checkpointBase: MiniMax H3Licence: not stated

Note: No image: none of the v4 preview images is rated PG on Civitai. The author's feature list (quality, audio, motion) is not independently tested.

Source: Civitai model page

Image: huggingface.co
Models

Qwen-Image-2.1-Turbo: 8-step generation and editing

Qwen released an accelerated checkpoint of Qwen-Image-2.1 for text-to-image and image editing in 8 denoising steps. It keeps the same 7B architecture, loads with QwenImage21Pipeline in Diffusers, ships its own sampling schedule and uses CFG 1 by default.

CheckpointBase: Qwen-Image-2.1Licence: Qwen Research License

Note: The Qwen Research licence is not a standard open-source licence: check its terms before commercial use.

Source: Hugging Face model card

Image: huggingface.co
Models

Veda brings learned sparse attention to MiniMax-H3 reference-to-video

Veda is a learned sparse-attention predictor for video diffusion. This preview applies it to MiniMax-H3 reference-to-audio-video generation, with the H3 Turbo LoRA for 8 steps. It is not a LoRA and leaves the backbone untouched; the predictor is stored in FP8.

Video add-onBase: MiniMax-H3Licence: MiniMax H3 Community License

Note: Preview release, fine-tuned for only 600 updates according to the card.

Source: Hugging Face model card

Image: blog.comfy.org
ComfyUI

Comfy Agent launches in Comfy Cloud

An in-app agent that plans, builds, runs and fixes ComfyUI workflows from natural-language requests. You can try it with Comfy Credits, and Comfy says it is coming to Comfy Desktop "in a few weeks".

Note: Not yet confirmed: the Comfy Desktop availability had not been announced as of 2026-10-26.

Source: Comfy blog

Image: civitai.com
civitai.red

Background: what civitai.red is

Since April 15, 2026, civitai.com is Civitai's SFW site (card payments, memberships, Creator Program), and civitai.red is the "freedom-first" site that shows all content, including NSFW, and takes crypto payments. Both sites share the same accounts and content, and civitai.green 301-redirects to civitai.com.

Note: Background explainer from April 2026, outside the current news window. The redirect from civitai.green to civitai.com was rechecked on 2026-10-26.

Source: Civitai article