HunyuanVideo 1.5 720p I2V ComfyUI template (video_hunyuan_video_1.5_720p_i2v): seven files, 20-step defaults and the bypassed 1080p upscale group

HunyuanVideo 1.5 720p image-to-video in ComfyUI: set up and run the official template Scope. This guide reads the official video_hunyuan_video_1.5_720p_i2v.json in Comfy-Org/workflow_templates at commit 0cc0fba562efe9f68e48050c2b82ed6495050bc2 (SHA256 8a86154c...7e33, 47,248 bytes). The graph is flat: no subgraph, no switch nodes. Optional stages are turned on or off by bypassing nodes (Ctrl+B), and in the saved file every optional stage is bypassed. Nothing was run and no weights were downloaded. Every value below is what the file saves or wires, not an observed result. 1. Prerequisites - A ComfyUI build with the HunyuanVideo 1.5 core nodes. All nodes in the file are comfy-core, with saved versions 0.3.68 to 0.3.71 and frontend 1.30.6. These are saved strings, not a tested minimum; update ComfyUI if a HunyuanVideo15 node is missing. No custom node pack is needed. - Five files for the default 720p path. The template's Model links note names the repos Comfy-Org/HunyuanVideo_1.5_repackaged and Comfy-Org/sigclip_vision_384, and gives these sizes: models/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors (8.74 GB) models/text_encoders/byt5_small_glyphxl_fp16.safetensors (418.3 MB) models/clip_vision/sigclip_vision_patch14_384.safetensors (816.8 MB) models/diffusion_models/hunyuanvideo1.5_720p_i2v_fp16.safetensors (15.51 GB) models/vae/hunyuanvideo15_vae_fp16.safetensors (2.35 GB) - Two more files only if you enable the 1080p group. The bypassed loaders declare them, but the note's list and storage tree leave them out, and the file gives no sizes: models/diffusion_models/hunyuanvideo1.5_1080p_sr_distilled_fp16.safetensors models/latent_upscale_models/hunyuanvideo15_latent_upsampler_1080p.safetensors 2. Inputs and settings on the default path - Start image: LoadImage (node 80) feeds HunyuanVideo15ImageToVideo (78) as start_image, and CLIPVisionEncode (79, crop "center") with the sigclip vision model (81). - Prompt: positive CLIPTextEncode (44) and an empty negative CLIPTextEncode (93), both on DualCLIPLoader (11) with qwen_2.5_vl_7b_fp8_scaled plus byt5_small_glyphxl_fp16, type hunyuan_video_15. - Size and length: node 78 saves 1280, 720, 121, 1, which in its widget order are width, height, frame length and batch. The file does not label them, and none are linked, so edit them on node 78 (the "Step 4 - Video Size" group). At the template's 24 fps, 121 frames is about 5 seconds by arithmetic. - Sampling: one SamplerCustomAdvanced (125) with RandomNoise (127, seed 887963123424675, fixed), KSamplerSelect euler, BasicScheduler (126: simple, 20 steps, denoise 1) and CFGGuider (129) at CFG 6 on ModelSamplingSD3 (130) at shift 7. The model comes from UNETLoader (12, hunyuanvideo1.5_720p_i2v_fp16). - Output: VAEDecode (8), then CreateVideo (101) at 24 fps, then SaveVideo (102) with prefix video/hunyuan_video_1.5, format auto, codec h264. 3. Steps versus the Hunyuan team settings The template's Sampler settings note explains the defaults: 50 inference steps take too long, so the template uses 20. Its table of original Hunyuan team settings lists 720p_i2v at CFG 6, shift 7 and 50 steps; the distilled variants use CFG 1. This template's CFG 6 and shift 7 already match the 720p_i2v row. For the original setting, raise BasicScheduler steps to 50. The 720p file here is not a distilled checkpoint, so the CFG 1 rows do not apply to it. 4. Run it 1) Put the five files in the folders above. 2) Open the template. 3) Load your start image in node 80. 4) Write the prompt in node 44. 5) Set width, height and length on node 78 if needed. 6) Run. Expect one video at video/hunyuan_video_1.5 (h264, 24 fps); with the saved values that is 121 frames at 1280x720 by arithmetic. 5. Optional stages (all bypassed in the saved file) - EasyCache (105: reuse_threshold 0.2, start 0.15, end 0.95) sits between the 720p UNET and both the scheduler and the shift node. A template note says it speeds up the workflow but lowers video quality; select it and press Ctrl+B to enable it. - Tiled decoding: VAEDecodeTiled (120: 512, 64, 64, 4096) takes the base sampler's latent, but its image output is connected to nothing. A template note suggests tiled decoding when decoding is slow. To use it, enable node 120 and connect its output to CreateVideo 101 in place of VAEDecode 8; enabling it alone changes nothing. - 1080p group ("Video Upscale 1080P (Ctrl-B to enable)"): the base latent goes to HunyuanVideo15LatentUpscaleWithModel (109, saved bilinear, 1920, 1080, disabled) with the latent upsampler (108), then to HunyuanVideo15SuperResolution (113, saved 0.7; it also takes the start image, CLIP vision output, VAE and text conditioning). An 8-step simple schedule (134) is split at step 4 (135). The first half runs sampler 138 with RandomNoise 889 (fixed) and CFGGuider 141 at CFG 1, on the 1080p SR distilled UNET (111) through ModelSamplingSD3 133 at shift 2. The second half runs sampler 139 without added noise and CFGGuider 140 at CFG 1. The group writes a second video at video/hunyuan_video_1.5_sr; the 720p SaveVideo stays active, so you get two files. 6. Failure cases and checks - Enabling the 1080p group without the two extra files: they are not in the note's download list. - Out of memory: a template note suggests setting weight_dtype to "fp8_e4n3fn". The FP8 format is e4m3fn, so look for the fp8_e4m3fn entry in UNETLoader's weight_dtype menu. The menu itself is not in this file. The file gives no VRAM figure. - Tiled decode not taking effect: node 120's output must be wired, as above. - Wiring to check before changing shift: in the base path, BasicScheduler 126 receives the model from EasyCache 105 (bypassed, so effectively the UNET), not from the shift-7 node 130 that feeds the guider. In the 1080p group, guider 140 takes the model directly from EasyCache 116 (not through shift node 133) and uses the plain text conditioning (44/93) rather than the super-resolution conditioning. The second tiled decoder (112) is connected to the SR sampler but its output is unused. The file does not say whether these differences are intended, and nothing was run to test their effect. - Expecting the Hunyuan team's quality at 20 steps: the template note itself says 50 steps is the original setting. 7. Related records "HunyuanVideo 1.5 720p T2V in diffusers: attention backend and default settings" covers the separate diffusers text-to-video example (50 steps, 121 frames, 24 fps). A separate README record covers memory and offloading statements for HunyuanVideo and HunyuanVideo-1.5. This guide covers the ComfyUI image-to-video template and its optional ComfyUI stages. Evidence and rights. Evidence status: source_inspected, not_run, unreviewed. This is shared-founder desk curation by the tutorial lane. The workflow repository declares the MIT licence at this commit; that covers this analysis of the template, not the model weights, encoders, VAE, upsampler, input images or generated videos, which keep their own terms.

{
  "citations": [
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/templates/video_hunyuan_video_1.5_720p_i2v.json",
      "note": "Exact pinned template; node byte ranges and excerpt hashes are in the source pins."
    },
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/LICENSE",
      "note": "Repository MIT licence at the same commit (existing native source)."
    }
  ]
}