How to run HunyuanVideo 1.5 locally.

480p or 720p.

Tencent released HunyuanVideo 1.5 on 20 November 2025: an 8.3B model that makes five-second clips at 24 fps, from text or from a picture. Every model is its own file, trained for one resolution and one task.

• 480p text to video and image to video. The ones to start with on 12 to 16 GB.
• 720p text to video and image to video. What Comfy’s templates use. Sharper, and much slower.
• CFG-distilled versions of both run at CFG 1, one pass per step instead of two. Tencent still asks for 50 steps.
• 480p I2V step-distilled, added on 5 December, runs in 8 or 12 steps. Tencent says an RTX 4090 makes a clip with it in about 75 seconds.
• Super-resolution models take a finished clip up to 720p or 1080p. More on that under settings.

Image to video needs one extra file, a SigLIP vision encoder. Text in the video, like a sign or a title, comes from the second text encoder, ByT5.

Files you need.

All from Comfy-Org’s repackage. The encoders and the VAE are shared by every model.

Shared parts

• Text encoder qwen_2.5_vl_7b_fp8_scaled.safetensors ComfyUI/models/text_encoders/ or qwen_2.5_vl_7b.safetensors, bf16, 16.6 GB
  9.4 GB Download
• Glyph encoder byt5_small_glyphxl_fp16.safetensors ComfyUI/models/text_encoders/
  0.4 GB Download
• VAE hunyuanvideo15_vae_fp16.safetensors ComfyUI/models/vae/
  2.5 GB Download

Text to video, pick one

• 480p, CFG-distilled hunyuanvideo1.5_480p_t2v_cfg_distilled_fp8_scaled.safetensors ComfyUI/models/diffusion_models/ fp16: 16.7 GB
  8.3 GB Download
• 480p hunyuanvideo1.5_480p_t2v_fp16.safetensors ComfyUI/models/diffusion_models/
  16.7 GB Download
• 720p hunyuanvideo1.5_720p_t2v_fp16.safetensors ComfyUI/models/diffusion_models/ the Comfy template’s model
  16.7 GB Download

Image to video, pick one, plus the vision file

• 480p, step-distilled hunyuanvideo1.5_480p_i2v_step_distilled_fp8_scaled.safetensors ComfyUI/models/diffusion_models/ fp16: 16.7 GB
  8.3 GB Download
• 720p hunyuanvideo1.5_720p_i2v_fp16.safetensors ComfyUI/models/diffusion_models/ the Comfy template’s model. CFG-distilled fp8: 8.3 GB
  16.7 GB Download
• Vision encoder sigclip_vision_patch14_384.safetensors ComfyUI/models/clip_vision/
  0.9 GB Download

Optional

• 1080p upscaler hunyuanvideo1.5_1080p_sr_distilled_fp8_scaled.safetensors ComfyUI/models/diffusion_models/ fp16: 16.7 GB. Needs the latent upsampler below
  8.3 GB Download
• Latent upsampler hunyuanvideo15_latent_upsampler_1080p.safetensors ComfyUI/models/latent_upscale_models/
  0.2 GB Download
• 4-step LoRA hunyuanvideo1.5_t2v_480p_lightx2v_4step_lora_rank_32_bf16.safetensors ComfyUI/models/loras/ for 480p text to video
  0.3 GB Download

models/
├── clip_vision/
│   └── sigclip_vision_patch14_384.safetensors   (image to video)
├── diffusion_models/
│   └── hunyuanvideo1.5_480p_t2v_cfg_distilled_fp8_scaled.safetensors
├── text_encoders/
│   ├── qwen_2.5_vl_7b_fp8_scaled.safetensors
│   └── byt5_small_glyphxl_fp16.safetensors
└── vae/
    └── hunyuanvideo15_vae_fp16.safetensors

The plain 720p models have no fp8 file. If they run out of memory, Comfy’s template note says to set weight_dtype in Load Diffusion Model to fp8_e4m3fn.

Smaller files: GGUF

jayn7’s GGUF builds load through ComfyUI-GGUF with Unet Loader (GGUF). The same uploader has 720p text-to-video and image-to-video GGUFs at similar sizes.

The repo has the plain and the CFG-distilled 480p model. Q8_0 is close to fp16.

What fits your computer.

The model file is only part of it. At 121 frames the VAE decode at the end is where 12 GB cards run out, and VAE Decode (Tiled) is the fix.

• 6 to 8 GBUntested

  No ComfyUI report found at this size. The 480p Q4_K_M GGUF is 5.1 GB. WanGP lists HunyuanVideo 1.5 from 6 GB.

• 12 GBOffloads

  480p in fp8 or GGUF. An RTX 3060 ran 480p at 121 frames. An RTX 4070 was fine at 81 frames and spilled over at 121, in the VAE decode.

• 16 GB480p fits

  480p fp8. An RTX 4060 Ti 16 GB ran the CFG-distilled 480p model at 640 × 480, 81 frames.

• 24 GBFits

  720p fp16. The default 720p text-to-video template took about 20 minutes on an RTX 4090.

• 32 GBFits

  RTX 5090: 848 × 480 at 121 frames in 284 s with the 720p model.

• Mac 48 GB+Untested

  No verified Mac report. fp8 files don’t load on Apple GPUs; the fp16 model, bf16 encoder and VAE add up to 35.8 GB.

• 12 GB graphics cards
• RTX 3060 12 GB
• 16 GB cards
• RTX 5060 Ti 16 GB
• 24 GB cards
• RTX 5090
• Mac with Apple Silicon

Set it up.

1. 1

   Update ComfyUI

   HunyuanVideo 1.5 needs core nodes from late November 2025, such as HunyuanVideo15ImageToVideo. ComfyUI Desktop was behind at first, so check that it has updated. In a manual install:

   Terminal, in the ComfyUI folder Copy

   git pull
   pip install -r requirements.txt

2. 2

   Download the files

   The two text encoders, the VAE and one model from the lists above. For image to video, add the SigLIP vision file.

3. 3

   Put them in their folders

   Model in diffusion_models, both encoders in text_encoders, VAE in vae, SigLIP in clip_vision. Restart ComfyUI.

4. 4

   Open the template

   In the template browser, pick Hunyuan Video 1.5 Text to Video or Hunyuan Video 1.5 Image to Video. Both are set up for the 720p models, with a 1080p upscale stage that starts switched off.

5. 5

   Check the loaders

   DualCLIPLoader gets Qwen2.5-VL and ByT5 with type hunyuan_video_15. For image to video, Load CLIP Vision gets the SigLIP file. If you swapped in a 480p model, set the size to 848 × 480 and the shift to 5.

6. 6

   Write a prompt and run

   Be specific about subject, motion and camera. Tencent’s own code rewrites prompts with a large language model before generating; in ComfyUI you write the full prompt yourself.

Settings that work.

Tencent sets the shift per model, and it matters. Comfy’s templates use 20 steps instead of 50, because 50 “just take too long”, and shift 7 on everything.

480p, Tencent’s settings

Steps

50Comfy template: 20

CFG

61 for CFG-distilled

Sampler

euler

Scheduler

simple

Shift

5Comfy template: 7

Size

848 × 480

Frames

1215 s at 24 fps

Negative

emptyTencent’s default

From Tencent’s model card, repeated in Comfy’s template notes. Comfy’s 720p templates use shift 7 for text to video too.

The 4-step lightx2v LoRA is the fast route for 480p text to video, at CFG 1. It also seems to work at 720p, but people found the results faded and low in contrast, especially in low light.

How fast.

Tencent’s 75 s figure is “within 75 seconds” for the step-distilled model. On the 2080 Ti, the VAE alone took about 20 s, and Wan 2.2 in a rapid merge took 31 s for the same clip.

When it goes wrong.

The video is black

SageAttention on an RTX 30-series card. Start ComfyUI without --use-sage-attention for this model.

Out of memory, or a crawl, at 121 frames on 12 GB

The VAE decode. Use VAE Decode (Tiled), use fewer frames, and on Windows set “Prefer No Sysmem Fallback” in the NVIDIA settings.

Expected tensor to have size 98 at dimension 1, but got size 64

A super-resolution file sits in the main model loader. Load a t2v or i2v model there; the SR model belongs in the upscale stage.

Node HunyuanVideo15ImageToVideo is missing

ComfyUI is too old. Update it; the Desktop app got the node later than the portable build.

Out of memory while loading the 720p model

Set weight_dtype in Load Diffusion Model to fp8_e4m3fn, or use a 480p fp8 file.

A CFG-distilled model looks unfinished

Tencent says the CFG-distilled models must use 50 steps. 20 is too few for them.

Faded, low-contrast video with the 4-step LoRA

A known trade-off of that LoRA. Use more steps without it, or the 480p step-distilled model for image to video.

Questions.

Sources: HunyuanVideo 1.5 model card, licence, Comfy-Org repackage, Comfy-Org workflow templates, generation times thread [1], repackage test thread [2], step-distilled thread [3], ComfyUI issue #10888 [4], 12 GB thread, #10855, ComfyUI issue #10823, SR model thread.