Model compatibility check

Double-check that all your loaded models (wan2.1_2v_720p_14B_fp8_e4m3fn.safetensors, c_umt5_xxl_fp8_e4m3fn_scaled.safetensors, vae_hunyuan_video_vae_bf16.safetensors) are compatible with each other and specifically intended for video generation with the architecture you're using (which seems to be based on Stability Diffusion given "Model/SamplingSD3" and "KSampler").

Below is a quick‐check checklist you can run through before you queue another job. (The file names in **bold** are the ones you said you have on disk.) --- ### 1. Are the three checkpoints drawn from the *same* video-model family? | File | Belongs to | What it is | |------|------------|------------| | **wan2.1_i2v_720p_14B_fp8_e4m3fn.safetensors**<br>(16 GB) | **Wan 2.1** | 14 B *image-to-video* diffusion backbone trained for 720 p output citeturn0search0 | | **c_umt5_xxl_fp8_e4m3fn_scaled.safetensors**<br>(6.7 GB) | **Wan 2.1** | UMT-5 XXL text encoder that pairs with *all* Wan 2.1 variants citeturn0search1 | | **hunyuan_video_vae_bf16.safetensors**<br>(493 MB) | **HunyuanVideo** | 3-D VAE trained by Tencent for the Hunyuan family citeturn0search2 | The first two files are a matched pair; the third comes from a *different* project. Wan 2.1 ships with its own 2-D VAE (`wan_2.1_vae.safetensors`, 254 MB) and that is what the official ComfyUI workflows load by default citeturn8view0turn9search0. **Mixing the Hunyuan VAE with Wan 2.1 usually works (the latent shapes match) but costs ~2-3× more VRAM and decode time and can shift colour balance**. If you see any of: * brown/green colour casts * very soft details / ringing * VAE-Decode-Tiled node taking >50 s per frame …switch to the native Wan VAE or to a lightweight replacement such as TAEHV’s Wan preset citeturn10search4. --- ### 2. Is your sampler/video pipeline expecting a **video** latent grid? ComfyUI’s generic **KSampler** node is image-centric; for Wan 2.1 you should be using either * `WanVideoKSampler` (community node) or * the built-in “Wan 2.1” example workflow (which connects `Model/SamplingSD3` → `KSampler` → `Decode Video`) Both variants ensure that the sampler runs **3-D denoising** and that the latent shape stays `(B, F, C, H, W)` instead of flattening the time dimension citeturn11search3. A plain KSampler will silently drop the extra frames and you end up with a single still image. --- ### 3. Do your *precision* and *hardware* line up? * `fp8_e4m3fn` checkpoints **require** NVIDIA Hopper (H100/H200) or Ada‐Lovelace cards with CUDA 12 + torch-compile. If you are on a consumer 30-series or earlier, convert (or download) the **fp16** or **bf16** versions instead; they’re drop-in replacements citeturn8view0. * Mixed precisions work fine as long as everything is ≥ the VAE precision. Example: fp8 diffusion + fp8 text encoder + bf16 VAE is OK; fp16 diffusion + fp8 VAE will crash. --- ### 4. Quick code probe (optional) ```python from safetensors.torch import load_file for f in ["wan2.1_i2v_720p_14B_fp8_e4m3fn.safetensors", "c_umt5_xxl_fp8_e4m3fn_scaled.safetensors", "wan_2.1_vae.safetensors"]: W = load_file(f, device="cpu", safe_metadata_only=True) print(f, W.metadata["__class__"]) ``` *All three should report a `"transformer"`-based class for the first two and `"AutoencoderKL"` (or similar) for the VAE. If the VAE reports `"Hunyuan3DVAE"` you’ve still got the Hunyuan file on the path.* --- ### What to do right now 1. **Download / swap in the native Wan VAE** (`wan_2.1_vae.safetensors`) and point your *Load VAE* node to it. 2. Update to the latest ComfyUI nightly (≥ 2025-05-20) so the bundled Wan example flows appear under *Examples → Video → Wan 2.1*. 3. If you must stay FP8, confirm your driver is CUDA 12.4+ and `torch-compile --mode=max-autotune` is enabled; otherwise pick the `*_fp16.safetensors` variants. 4. Use a Wan-aware sampler node (or the reference workflow) rather than a plain KSampler. Once those four boxes are ticked, all three checkpoints will line up with the SD3-style video architecture and you should be able to render 720 p clips without shape or precision errors.

pick the *_fp16.safetensors variants.

Here's how to proceed if you're switching to the `*_fp16.safetensors` versions of the Wan 2.1 video generation model: --- ### ✅ Recommended File Set (FP16 versions): | **Component** | **File Name** | **Size** | **Required?** | |---------------|----------------|----------|----------------| | Diffusion backbone | `wan2.1_i2v_720p_14B_fp16.safetensors` | ~26 GB | ✅ Required | | Text encoder | `c_umt5_xxl_fp16_scaled.safetensors` | ~11 GB | ✅ Required | | VAE (official) | `wan_2.1_vae.safetensors` | ~254 MB | ✅ Required | | VAE (alt option) | `taehv_vae_720p_wa.safetensors` | ~110 MB | Optional (smaller, faster) | --- ### 🔧 Where to get them You can usually find these on: - [CivitAI](https://civitai.com/models/)—search for `wan2.1_i2v_fp16` - [Hugging Face](https://huggingface.co/)—e.g., [`ali-vilab/Wan2.1`](https://huggingface.co/ali-vilab/wan2.1) - GitHub releases from authors like Ali-VILab, TencentARC, or community repackers Make sure you're **not downloading the `bf16` or `fp8` versions** by mistake. --- ### 🛠️ In ComfyUI Once downloaded: 1. **Move files into your `models/checkpoints/` and `models/vae/` directories.** 2. **In your ComfyUI flow:** - Use a **`Load Checkpoint`** node pointing to `wan2.1_i2v_720p_14B_fp16.safetensors` - Use a **`Load Text Encoder`** or `Load CLIP` (if Wan2.1 flow expects it) for `c_umt5_xxl_fp16_scaled.safetensors` - Use a **`Load VAE`** node for `wan_2.1_vae.safetensors` 3. Make sure you use a **Wan-compatible sampler**, e.g. `ModelSamplingSD3` → `KSamplerAdvanced` → `DecodeVideo` --- ### 🧠 Why FP16? The `fp16` versions are: - Compatible with most GPUs (RTX 3060/3070/3080/3090, A6000, etc.) - Much more stable than `fp8` which requires Hopper GPUs or complex optimization - Still VRAM-efficient and performant for batch rendering --- ### 🚨 Caution - If you're mixing with other models (like SDXL or AnimateDiff), make sure the samplers and VAEs are not clashing. - Do **not** use the original Hunyuan 3D VAE unless you're working with a Hunyuan-trained video model. It's incompatible with Wan-style latent shapes in subtle ways. --- Let me know your GPU specs if you’d like an optimized setup (e.g., max resolution per frame, batch size, etc.).

no I want to do t2v

Perfect — if you're aiming for **text-to-video (T2V)** with **Wan 2.1** at **FP16 precision**, here's the full setup you need to run it properly in **ComfyUI** (or similar SD3-compatible architecture). --- ## ✅ Components You Need (All FP16 T2V-Compatible) | Component | File | Notes | |----------|------|-------| | 🎥 **Video diffusion model** | `wan2.1_t2v_720p_14B_fp16.safetensors` | The *T2V-specific* 720p version of Wan 2.1 in FP16 precision (NOT `i2v`) | | 🧠 **Text encoder** | `c_umt5_xxl_fp16_scaled.safetensors` | Paired with all Wan 2.1 T2V variants (UMT5-XXL) | | 🌀 **VAE** | `wan_2.1_vae.safetensors` | Use the official VAE for Wan (2D, fast, color-accurate) | | 🛠️ Optional VAE | `taehv_vae_720p_wa.safetensors` | Optional lighter alternative (faster, slightly less fidelity) | --- ### ❌ Do NOT use: - `wan2.1_i2v_720p_14B_fp16.safetensors` ← this is for *image-to-video*, not text-to-video - Hunyuan VAE (`hunyuan_video_vae_bf16.safetensors`) ← mismatched latent format - `fp8` versions unless you’re on Hopper GPUs (H100, L40S, etc.) --- ## 🧩 Architecture Flow in ComfyUI (T2V) If you're using a ComfyUI workflow, the high-level flow looks like: ``` 1. Load Text Prompt ↓ 2. Load Checkpoint (`wan2.1_t2v_720p_14B_fp16.safetensors`) ↓ 3. Load UMT5 XXL Text Encoder (`c_umt5_xxl_fp16_scaled.safetensors`) ↓ 4. KSampler / Wan-Specific Sampler (with temporal denoising) ↓ 5. Decode latent → RGB using `wan_2.1_vae.safetensors` ↓ 6. Save Video / Preview ``` Make sure your sampler **preserves the time dimension (F)** — Wan uses 3D latents of shape `(B, F, C, H, W)`. --- ## 🔌 Download Sources - **Hugging Face**: [`ali-vilab/Wan2.1`](https://huggingface.co/ali-vilab/wan2.1) - **CivitAI**: search `wan2.1_t2v_fp16` - Or use model indexers like [civitai.com](https://civitai.com), [modelscope.cn](https://modelscope.cn), or direct links from GitHub releases (e.g., `WanT2VLoader`) --- ## 🚀 Pro Tips for Better Results - Stick to **720p (1280×720)** for full support. You can upscale later. - Use **prompt + negative prompt** to guide coherence. - Try **10–16 frames** per sample to keep VRAM under control unless you're on 24+ GB cards. - If using ComfyUI, grab the [official Wan 2.1 T2V example](https://github.com/comfyanonymous/ComfyUI_experiments) from the *Examples → Video → Wan 2.1 Text-to-Video* section in the nightly release (post-May 2024). --- Want a pre-built `.json` ComfyUI workflow file for T2V with these settings? Or a guide to optimize for your GPU? Just ask.