RedCraft-红潮 REDMixBaked

做好工具人 服务艺术家
Forever in memory of METAFILM Studio founder Mr. Yuan Bo

8-21 REDGraft LTX 2.5 老同学 | sulphur2 ported version 高速移植版 Fast 2K
Sulphur 2 Base is an uncensored, community-built video generation model on top of LTX 2.3. It runs text-to-video and image-to-video natively and supports every other LTX 2.3 format, with an optional prompt enhancer and distill LoRA in the same release.

Originally released by SulphurAI on Hugging Face. All credit for the model goes to the SulphurAI team and contributors below. Civitai is hosting a mirror so creators can run it on-site - please head to the original repo for weights, updates, and to support the project directly.

Built by

• SulphurAI - Hugging Face | Discord | Ko-fi

• TenStrip - testing and merging (LTX2.3-10Eros i2v merge)

• silveroxides (@s1lv3rc01n) - testing, merging, quantizing

• @mov7162 - Musubi Tuner guidance

• Two anonymous funders who made the original Sulphur and Sulphur 2 possible

• Many other community contributors

Same-Architecture Model Transfer
同源同构模型权重移植方法

Structured Weight-Space Transfer Between Same-Family Diffusion Models

Technical write-up of a structured weight-transfer methodology for moving learned model characteristics between checkpoints that share the same model family, architecture, tensor layout, and inference structure.

The method is designed for situations where a source checkpoint and target checkpoint are architecturally identical or near-identical, allowing their corresponding parameter tensors, attention heads, MLP projections, and transformer blocks to be directly aligned.

Rather than treating checkpoint merging as a blind weighted average, this methodology treats the difference between two trained checkpoints as a structured parameter subspace and selectively transfers that subspace into the target model.

can be aligned directly without dimensional projection.

This makes it possible to perform much more precise parameter-space operations than cross-architecture grafting.

The transfer can operate at several granularities:

• transformer block

• attention head

• Q / K / V / O projection

• MLP up / gate / down projection

• individual tensor bands

• selected block ranges

• weighted parameter subspaces

The objective is not to replace the target model with the source model.

The objective is to inject selected learned characteristics from the source while preserving the target model's existing functional structure.

Why Same-Architecture Transfer Is Fundamentally Different

Cross-architecture transfer requires solving several alignment problems，but

If:

WS(l)​∈Rm×n

and:

WT(l)​∈Rm×n

represent the same tensor at the same layer, then:

WS(l)​↔WT(l)​

is a direct correspondence.

No dimensional projection is required.

No head remapping is required.

No truncation is required.

No GQA conversion is required.

This means the source and target models occupy the same parameter coordinate system.

Therefore:

• Q influences query formation

• K influences attention addressing

• V carries the attended information

• O projects the resulting attention representation

A structured transfer can therefore test each component independently.

The content above is generated by GPT and represents a "CivitAI-science" style of radical reinterpretation (Just for fun，not formal scientific research); it is intended solely for the purpose of popularizing basic knowledge.

同架构模型权重移植

—— Same-Architecture Structured Weight Transfer

这是一种针对同源、同架构模型的结构化权重移植方法。与传统的模型平均（Model Merge）不同，它不追求简单地混合两个 checkpoint，而是尝试从源模型中提取特定的参数变化方向 / 功能子空间，再将其有选择地注入目标模型。

权重方向迁移

除了直接：WT′​=(1−α)WT​+αWS​

还可以提取源模型的参数变化：ΔW=WS​−Wbase​

然后：WT′​=WT​+αΔW

这种方式更接近 Task Vector / Parameter Delta 的思想。

它表达的不是：“把源模型复制一部分给目标模型。”

而是：“把源模型训练过程中形成的某种参数变化注入目标模型。”

一句话总结

同架构模型移植的核心，不是简单的 checkpoint Merge，而是利用源模型与目标模型处于同一参数空间这一优势，将源模型形成的参数差异、方向或特定子空间进行选择性提取，再以可控强度注入目标模型。

它可以理解为：

源模型 ↓

提取 Parameter Delta ↓

选择 Block / Head / QKV / MLP ↓

归一化 / 正交化 / 加权 ↓

注入目标模型 ↓

获得新的功能组合模型

相比跨架构 Graft，这种方法的理论基础更加直接，也更适合进一步研究 Task Vector、Model Merging、Subspace Transfer 和 Diffusion Transformer 功能迁移。

以上内容来自GPT，民科魔改向研究（纯娱乐，非科研），仅为普及基础知识。

1秒5帧！25 frames in 5 seconds
LTXV 2.3 GTAnimation 高速版
KREA 2 赤佬 Bastard3 Edition
REDZ 2 红潮 造相2 HDEdition

INT8/INT4 Convrot for ComfyUI 0.27 [Native 原生节点支持] Uploaded
File name: REDGTA1.0_LTX23_ComfyUI-int4_convrot(on this page)
File name: Krea2RedMix1.1-INT8-Convrot-ComfyUI (original native)
File name: Krea2RedMix2.1-INT8-Convrot-ComfyUI (no mosaics)
File name: Krea2RedMix3.1-INT8/INT4-Convrot-ComfyUI (visual effects)
File name: REDZimageTurbo2.0-INT8-Convrot-ComfyUI (ZIT-HD 2026)
File name: Krea2RedMix1.2-INT4-Convrot-ComfyUI (by Wikee Yang)

Sample WF（ComfyUI native node support）：https://civitai.red/models/579280

GPU with 22GB of VRAM can fully load the model and generate it at high speed;
GPU with 12GB or more of VRAM can generate it using weight swapping.

All in One, Wan for All

We are excited to introduce our latest model to our talented community creators:

Wan2.1-VACE, All-in-One Video Creation and Editing model.

Model size: 1.3B, 14B License: Apache-2.0

If we are in Wan Day, what will it be like? 如果我们在万相世界，会是什么样子？

模型支持两种文本到视频模型（1.3B 和 14B）和两种分辨率（480P 和 720P）。

WAN-VACE is not a T2V model per se, but rather R(reference)2V, Can be understood as Video ControlNet for WAN , so there is no way to provide a T2V workflow. The CausVid accelerator is a distillation accelerator technology that can be used on WAN-VACE to provide 4-8 steps of accelerated generation.

WAN-VACE本身不是T2V模型，而是R（参考）2V，可以理解为WAN的视频CN，因此无法提供T2V工作流程。CausVid加速器是一种蒸馏加速技术，可用于WAN-VACE，提供4-8步加速生成。

Introduction

VACE is an all-in-one model designed for video creation and editing. It encompasses various tasks, including reference-to-video generation (R2V), video-to-video editing (V2V), and masked video-to-video editing (MV2V), allowing users to compose these tasks freely. This functionality enables users to explore diverse possibilities and streamlines their workflows effectively, offering a range of capabilities, such as Move-Anything, Swap-Anything, Reference-Anything, Expand-Anything, Animate-Anything, and more.

VACE是一款专为视频创建和编辑而设计的一体化模型。它包括各种任务，包括视频生成（R2V）、视频到视频编辑（V2V）和屏蔽视频到视频剪辑（MV2V），允许用户自由组合这些任务。此功能使用户能够探索各种可能性，并有效地简化他们的工作流程，提供一系列功能，如移动任何内容、交换任何内容、引用任何内容、扩展任何内容、为任何内容设置动画等。

About CausVid-Wan2-1:

5-16 The PERFECT solution to CausVid from Kijai （Best practices）

Wan21_CausVid_14B_T2V_lora_rank32.safetensors · Kijai/WanVideo_comfy

Through weight extraction and block separation,

KJ give us a universal CausVid LoRA in rank32 for Any 14B WAN model,

EVEN including FT models and I2V model！

Although this may not have been CausVid's initial intention, by flexibly adjusting the LoRA parameters (0.3~0.5), we have achieved unprecedented availability on home grade graphics cards！

KJ-Godlike also provides a 1.3B bidirectional inference version of LoRA export file

Wan21_CausVid_bidirect2_T2V_1_3B_lora_rank32.safetensors

same time, we also noticed that xunhuang1995 uploaded the Warp-4Step_cfg2 autoregressive version 1.3B CausVid model from: tianweiy/CausVid

相与为壹，全部在万

Best Adaptation for WAN-VACE full Models

5/15 REDCausVid-Wan2-1-14B-DMD2-FP8 Uploaded 8-15 steps CFG 1

本页面右侧下载列表，Safetensors 格式，workflow 在 Trainning data 压缩包内

The download list on the right side of this page is in Safetensors format, and the workflow is included in the Training data compressed file. The example images and videos also include workflows (yes, you can directly throw the original video files into ComfyUI and try to capture the workflow)

5/15 Aiwood WAN-ACE Fully functional workflow Uploaded

5/15 ComfyUI KJ-WanVideoWrapper have been updated

5/14 autoregressive_checkpoint.pt 1.3b Uploaded , PT UNET Loader

5/14 bidirectional_checkpoint2.pt 1.3b Uploaded , PT UNET Loader

NEW Sampler Flowmatch_causvid in KJ-WanVideoWrapper

Releases from：

kijai/ComfyUI-WanVideoWrapper

⭐ leave a star⭐

[ The adaptability test results of WAN1.2 LoRAs for VACE show that about 75% of I2V/T2V LoRA weights can take effect, but the sensitivity is reduced ( try to increase the LoRA weight ,more than 100% Sometimes it can be helpful ) ]

Fullview of Aiwood WAN-ACE Fully functional workflow：

source: https://www.bilibili.com/video/BV1FGE6zGEDK ⭐ leave a star⭐

CausVid 加速器项目页 https://causvid.github.io/

WAN-VACE 模型的参数和配置如下：

📌 Wan2.1-VACE provides solutions for various tasks, including reference-to-video generation (R2V), video-to-video editing (V2V), and masked video-to-video editing (MV2V), allowing creators to freely combine these capabilities to achieve complex tasks.

👉 Multimodal inputs enhancing the controllability of video generation.

👉 Unified single model for consistent solutions across tasks.

👉 Free combination of capabilities unlocking deeper creative

📌 Wan2.1-VACE为各种任务提供解决方案，包括参考视频生成（R2V）、视频到视频编辑（V2V）和屏蔽视频到视频剪辑（MV2V），允许创作者自由组合这些功能来实现复杂的任务。

👉 多模态输入增强了视频生成的可控性。

👉 统一的单一模型，实现跨任务的一致解决方案。

👉 自由组合功能，释放更深层次的创造力

WAN实时生成来了！Hybrid AI model crafts smooth, high-quality videos in seconds

The CausVid generative AI tool uses a diffusion model to teach an autoregressive (frame-by-frame) system to rapidly produce stable, high-resolution videos.

Wan2.1based 混合AI模型在几秒钟内（9帧/秒）制作出流畅、高质量的视频

CausVid生成AI工具使用扩散模型来指导自回归（逐帧）系统快速生成稳定的高分辨率视频。

Hybrid AI model crafts smooth, high-quality videos in seconds | MIT News | Massachusetts Institute of Technology

From Slow Bidirectional to
Fast Autoregressive Video Diffusion Models

CausVid https://causvid.github.io/

tianweiy (Tianwei Yin)

RedCaus/REDCausVid-Wan2-1-14B-DMD2-FP8 Uploaded / WAN-VACE14B 最佳适配

CausVid/autoregressive_checkpoint uploaded / 自回归模型基于 WAN1.3B 已收录

CausVid/bidirectional_checkpoint2 uploaded / 双向推导模型基于 WAN1.3B 已收录

Kijai/Wan2_1-T2V-14B_CausVid_fp8_e4m3fn.safetensors / HF仓库 WanVideo_comfy

⭐ leave a star⭐

licensed by Creative Commons Attribution Non Commercial 4.0

Thank you for this friend's additional comment. I was too excited last night and didn't sleep, so I stopped updating before finishing：

We’ll need to use the official Python-based inference codes

1) Clone https://github.com/tianweiy/CausVid and follow instructions to install requirements

2) Clone https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B into wan_models/Wan2.1-T2V-1.3B

3) Put the pt file inside checkpoint_folder/model.pt

4) Run inference code, python minimal_inference/autoregressive_inference.py --config_path configs/wan_causal_dmd.yaml --checkpoint_folder XXX --output_folder XXX --prompt_file_path XXX

Reddit posts about CausVid: https://www.reddit.com/r/StableDiffusion/comments/1khjy4o/causvid_generate_videos_in_seconds_not_minutes/

https://www.reddit.com/r/StableDiffusion/comments/1k0gxer/causvid_from_slow_bidirectional_to_fast/

We have tested the CausVid based on Wan1.3b version, which has incredible speed, and are currently testing the 14B version produced by lightx2v.

LightX2V: Light Video Generation Inference Framework

Supported Model List

✅ HunyuanVideo-T2V

✅ HunyuanVideo-I2V

✅ Wan2.1-T2V

✅ Wan2.1-I2V

✅ Wan2.1-T2V-CausVid

✅ SkyReels-V2-DF

How to Run

Please refer to the documentation in lightx2v.

⭐ leave a star⭐

通义实验室 WAN 2.1 Model Zoo

Institute for Intelligent Computing专注于各领域大模型技术研发与创新应用。实验室研究方向涵盖自然语言处理、多模态、视觉AIGC、语音等多个领域。我们并积极推进研究成果的产业化落地。实验室同时积极参与开源社区建设，全方位拥抱开源社区，共同探索AI模型的开源开放。

Developer / Models Name / Kijai`s ComfyUI Model

RedCaus/REDCausVid-Wan2-1-14B-DMD2-FP8 Uploaded / WAN-VACE14B 最佳适配

CausVid/autoregressive_checkpoint included / 自回归模型基于 WAN1.3B 已收录

CausVid/bidirectional_checkpoint2 included / 双向推导模型基于 WAN1.3B 已收录

CausVid/wan_causal_ode_checkpoint_model testing / 自回归因果推导 测试中

CausVid/wan_i2v_causal_ode_checkpoint_model testing / 文生图模型 测试中

lightx2v/Wan2.1-T2V-14B-CausVid unqualify / 自回归模型14B AiWood实测不达标

lightx2v/Wan2.1-T2V-14B-CausVid quant unqualify / 自回归模型14B量化版 实测不达标

Wan Team/1.3B text-to-video included / 文生视频1.3B 已收录

Wan Team/14B text-to-video included / 文生视频14B 已收录

Wan Team/14B image-to-video 480P included / 图生视频14B 已收录

Wan Team/14B image-to-video 720P included / 图生视频14B 已收录

Wan Team/14B first-last-frame-to-video 720P included / 视频首尾帧 已收录

Wan Team/Wan2_1_VAE included / KiJai‘s WAN视频VAE 已收录

ComfyORG/Wan2.1_VAE included / Comfy‘s WAN视频VAE 已收录

google/umt5-xxl umt5-xxl-enc safetensors included / TE编码器 已收录

mlf/open-clip-xlm-roberta-large-vit-huge-14 safetensors included / CLIP编码器 已收录

DiffSynth-Studio Team/1.3B aesthetics LoRA 美学蒸馏-通义万相2.1-1.3B-LoRA-v1

DiffSynth-Studio Team/1.3B Highres-fix LoRA 高分辨率修复-通义万相2.1-1.3B-LoRA-v1

DiffSynth-Studio Team/1.3B ExVideo LoRA 长度扩展-通义万相2.1-1.3B-LoRA-v1

DiffSynth-Studio Team/1.3B Speed Control adapter 速度控制-通义万相2.1-1.3B-适配器-v1

PAI Team/ WAN2.1 Fun 1.3B InP 支持首尾帧 / Kijai/WanVideo_comfy

PAI Team/ WAN2.1 Fun 14B InP 支持首尾帧 / Kijai/WanVideo_comfy

PAI Team/ WAN2.1 Fun 1.3B Control 控制器 / Kijai/WanVideo_comfy

PAI Team/ WAN2.1 Fun 14B Control 控制器 / Kijai/WanVideo_comfy

PAI Team/ WAN2.1 Fun 14B Control 控制器 / Kijai/WanVideo_comfy

PAI Team/ WAN2.1-Fun-V1_1-14B-Control-Camera / Kijai/WanVideo_comfy

IIC Team/ VACE-通义万相2.1-1.3B-Preview / Kijai/WanVideo_comfy

IC ( In-Context ) Controler 多模态控制器 :

ali-vilab/ VACE: All-in-One Video Creation and Editing / Kijai/WanVideo_comfy

Phantom-video/Phantom Subject-Consistent via Cross-Modal Alignment

KwaiVGI/ ReCamMaster Camera-Controlled 镜头多角度 / Kijai/WanVideo_comfy

Digital Character 数字人 via Wan2.1 :

ali-vilab/ UniAnimate-DiT 长序列骨骼角色视频 / Kijai/WanVideo_comfy

Fantasy-AMAP/ 音频驱动数字人 FantasyTalking / Kijai/WanVideo_comfy

Fantasy-AMAP/ 角色一致性身份保留 FantasyID / Fantasy-AMAP/fantasy-id

Uncensored NSFW 解锁版本：

REDCraft AIGC / WAN2.1 720P NSFW Unlocked / forPrivate use【非公开】

CubeyAI / WAN General NSFW model (FIXED) / The Best Universal LoRA

昆仑万维发布 SkyReels based on Wan2.1

Skywork / SkyReels-V2-I2V-14B-720P / Image-to-Video / Kijai/WanVideo_comfy

Skywork / SkyReels-V2-I2V-14B-540P / Image-to-Video / Kijai/WanVideo_comfy

Skywork / SkyReels-V2-T2V-14B-540P / Text-to-Video / Kijai/WanVideo_comfy

Skywork / SkyReels-V2-T2V-14B-720P /Text-to-Video / Kijai/WanVideo_comfy

Skywork / SkyReels-V2-I2V-1.3B-540P / Image-to-Video / Kijai/WanVideo_comfy

AutoRegressive Diffusion-Forcing 无限长度生成架构

Skywork / SkyReels-V2-DF-14B-720P / Text-to-Video / Kijai/WanVideo_comfy

Skywork / SkyReels-V2-DF-14B-540P / Text-to-Video / Kijai/WanVideo_comfy

Skywork / SkyReels-V2-DF-1.3B-540P / Text-to-Video / Kijai/WanVideo_comfy

昆仑万维发布 SkyReels 视频标注模型：

Skywork / SkyCaptioner-V1 Skywork (Skywork) / Skywork/SkyCaptioner-V1

Tiny AutoEncoder / taew2_1 safetensors / Kijai/WanVideo_comfy

A tiny distilled VAE model for encoding images into latents and decoding latent representations into images

WAN Comfy-Org/Wan_2.1_ComfyUI_repackaged

【例图页面蓝色Nodes或下载webp文件-可复现视频工作流】

Gallery sample images/videos (WEBP format) including the ComfyUI native workflow

This is a concise and clear GGUF model loading and tiled sampling workflow：

Wan 2.1 Low vram Comfy UI Workflow (GGUF) 4gb Vram - v1.1 | Wan Video Workflows | Civitai

节点：（或使用 comfyui manager 安装自定义节点）

https://github.com/city96/ComfyUI-GGUF

https://github.com/kijai/ComfyUI-WanVideoWrapper

https://github.com/BlenderNeko/ComfyUI_TiledKSampler

* 注意需要更新到最新版本的 comfyui-KJNodes GitHub - kijai/ComfyUI-KJNodes: Various custom nodes for ComfyUI update to the latest version of Comfyui KJNodes

Kijai ComfyUI wrapper nodes for WanVideo

WORK IN PROGRESS

@kijaidesign 's works

Huggingface - Kijai/WanVideo_comfy

GitHub - kijai/ComfyUI-WanVideoWrapper

主图视频来自 AiWood

https://www.bilibili.com/video/BV1TKP3eVEue

Text encoders to ComfyUI/models/text_encoders

Transformer to ComfyUI/models/diffusion_models

Vae to ComfyUI/models/vae

Right now I have only ran the I2V model succesfully.

Can't get frame counts under 81 to work, this was 512x512x81

~16GB used with 20/40 blocks offloaded

DiffSynth-Studio Inference GUI

Wan-Video LoRA & Finetune training.

DiffSynth-Studio/examples/wanvideo at main · modelscope/DiffSynth-Studio · GitHub

💜 Wan    ｜    🖥️ GitHub    |   🤗 Hugging Face   |   🤖 ModelScope   |    📑 Paper (Coming soon)    |    📑 Blog    |   💬 WeChat Group   |    📖 Discord  

Wan: Open and Advanced Large-Scale Video Generative Models

通义万相Wan2.1视频模型开源！视频生成模型新标杆，支持中文字效+高质量视频生成

In this repository, we present Wan2.1, a comprehensive and open suite of video foundation models that pushes the boundaries of video generation. Wan2.1 offers these key features:

• 👍 SOTA Performance: Wan2.1 consistently outperforms existing open-source models and state-of-the-art commercial solutions across multiple benchmarks.

• 👍 Supports Consumer-grade GPUs: The T2V-1.3B model requires only 8.19 GB VRAM, making it compatible with almost all consumer-grade GPUs. It can generate a 5-second 480P video on an RTX 4090 in about 4 minutes (without optimization techniques like quantization). Its performance is even comparable to some closed-source models.

• 👍 Multiple Tasks: Wan2.1 excels in Text-to-Video, Image-to-Video, Video Editing, Text-to-Image, and Video-to-Audio, advancing the field of video generation.

• 👍 Visual Text Generation: Wan2.1 is the first video model capable of generating both Chinese and English text, featuring robust text generation that enhances its practical applications.

• 👍 Powerful Video VAE: Wan-VAE delivers exceptional efficiency and performance, encoding and decoding 1080P videos of any length while preserving temporal information, making it an ideal foundation for video and image generation.

This repository features our T2V-14B model, which establishes a new SOTA performance benchmark among both open-source and closed-source models. It demonstrates exceptional capabilities in generating high-quality visuals with significant motion dynamics. It is also the only video model capable of producing both Chinese and English text and supports video generation at both 480P and 720P resolutions.