SuShuHeng/wan2.2_it2v_5B_Turbo_ConvRot_INT8_INT4

Wan2.2-TI2V-5B-Turbo ConvRot — INT4 / INT8 / BF16-F32 Mixed

ComfyUI-native, single-file diffusion-transformer conversions of the 4-step Wan2.2-TI2V-5B-Turbo model. This repository contains only the diffusion transformer. It does not include a VAE, UMT5 text encoder, tokenizer, scheduler, or complete Diffusers pipeline.

Included files

File	Format	Size	Static verification
wan2.2_ti2v_5b_turbo_int4_convrot.safetensors	ConvRot W4A4, group size 256	2.48 GiB	300 quantized layers; convrot_w4a4 metadata
wan2.2_ti2v_5b_turbo_int8_convrot.safetensors	INT8 Tensorwise ConvRot, group size 256	4.77 GiB	300 quantized layers; int8_tensorwise, convrot=true metadata
wan2.2_ti2v_5b_turbo_bf16_f32_mixed.safetensors	ComfyUI-native BF16/F32 mixed precision	9.33 GiB	825 tensors; no quantization metadata

All three files originate from the same converted source revision: 15f1944bf867c08e18dcf0a79e41a6176e590edc of yetter-ai/Wan2.2-TI2V-5B-Turbo-Diffusers.

ComfyUI use

1. Place one model file in ComfyUI/models/diffusion_models/.
2. Place the matching example start image from examples/ in ComfyUI/input/.
3. Load the matching workflow from workflows/ and select that image in its LoadImage node.
4. Obtain the dependency models referenced by the workflow separately: wan2.2_vae.safetensors and the chosen UMT5 text encoder. They are not redistributed here.

The supplied workflows preserve the Turbo operating point: 4 steps, CFG 1.0, UniPC BH2, simple scheduler, flow shift 5.0, 24 FPS, and 121 frames. They also use tiled VAE decode settings 736 / 96 / 128 / 16 (spatial tile / spatial overlap / temporal tile / temporal overlap) and encode through GJJ_VideoCombine.

Workflow	Diffusion model	New, original example scene
06_video_produce_turbo_lite.json	INT4 ConvRot	survey rover on a salt flat
06_video_produce_turbo_plus.json	INT8 ConvRot	paper sailboat on a mountain lake
06_video_produce_turbo_pro.json	BF16/F32 mixed	glass greenhouse on a hillside

The example prompts, negative prompts, and start images are newly created for this repository. They contain no people, brands, readable text, logos, or copyrighted characters. The three images are AI-generated illustrative starting frames, not benchmark evidence.

Provenance and license

This is a technical quantization and key-format conversion, not an official release and not a claim of endorsement by Wan-AI, quanhaol, or yetter-ai.

• Upstream base model: Wan-AI/Wan2.2-TI2V-5B, published under Apache-2.0.
• Turbo model converted here: yetter-ai/Wan2.2-TI2V-5B-Turbo-Diffusers, whose card identifies it as the Diffusers version of the Turbo model.
• Turbo source project and license: quanhaol/Wan2.2-TI2V-5B-Turbo and its CC BY-NC-SA 4.0 license.

The Turbo project's LICENSE.md is Creative Commons Attribution–NonCommercial–ShareAlike 4.0 International (CC BY-NC-SA 4.0). Quantized model weights are adapted material; therefore this repository applies the same CC-BY-NC-SA-4.0 license to the derived weight files and workflows. Attribution, a modification notice, the upstream links, and the license URL are provided here as required. In particular, do not use or redistribute these derived weights for commercial purposes, and share adaptations under compatible CC BY-NC-SA terms. This is a factual provenance notice, not legal advice.

Checksums

File	SHA-256
wan2.2_ti2v_5b_turbo_int4_convrot.safetensors	4136F7AC64810AA2696C285F8FCE69EA20E70112419139AB1E2EA1184DA55C4F
wan2.2_ti2v_5b_turbo_int8_convrot.safetensors	329CD5825F4580D455C52EBA5FCFE6F9CB608B492AE81503C709A8D6E003714D
wan2.2_ti2v_5b_turbo_bf16_f32_mixed.safetensors	6A84F30FE3C6332D09408B7293D247C7D541E9FF2B9BC93FA6247AC14761BC6E

Citation

Please cite the upstream Wan work when it is relevant to your project:

@article{wan2025,
  title={Wan: Open and Advanced Large-Scale Video Generative Models},
  author={Team Wan and Ang Wang and Baole Ai and others},
  journal={arXiv preprint arXiv:2503.20314},
  year={2025}
}