Nodes/Comfy_HunyuanImage3/Hunyuan Unified Generate V2
ComfyUI Node

Hunyuan Unified Generate V2

A ComfyUI node in Hunyuan/V2 with 15 inputs and 2 outputs.

By EricRollei·Created 9 months ago·Updated 3 months ago· 64
Hunyuan Unified Generate V2
    • images
    • final_prompt
    model_nameHunyuanImage-3-NF4
    prompta beautiful sunset over mountains
    resolution1024x1024 (1:1 Square)
    num_inference_steps40
    guidance_scale5.0
    seed-1
    blocks_to_swap20
    vae_placementauto
    post_actionfull_unload
    enable_vae_tilingfalse
    flow_shift2.8
    reserve_vram_gb0.0
    moe_drop_tokenstrue
    vae_dtypebfloat16
    force_reloadfalse
    CategoryHunyuan/V2

    Inputs (15)

    NameTypeDefaultDescription
    model_nameCOMBOHunyuanImage-3-NF4Model folder. Quant type is auto-detected from name (NF4/INT8/BF16).
    promptSTRINGa beautiful sunset over mountainsText prompt for image generation.
    resolutionCOMBO1024x1024 (1:1 Square)Image resolution at common photo ratios (~1MP base, ~1.5MP HD, ~2.4MP large). All divisible by 16.
    num_inference_stepsINT4010–100Number of diffusion steps. 40 is balanced for ~1MP. Higher (50–80) reduces flow-matching artifacts at 2K+ resolutions but generation time scales linearly — expect a much longer wait.
    guidance_scaleFLOAT5.01–20CFG scale. Higher = more prompt adherence. 5.0-7.0 typical.
    seedINT-1-1–2147483647-1 = random seed.
    blocks_to_swapoptINT20-1–31-1 = auto calculate. 0 = no swapping (NF4 needs ~50GB; BF16 uses device_map). 1-31 = manual swap count. BF16 with block swap loads to CPU and is much faster than device_map.
    vae_placementoptCOMBOautoauto: decide based on VRAM. always_gpu: VAE stays on GPU. managed: VAE moves to CPU when not decoding.
    post_actionoptCOMBOfull_unloadkeep_loaded: Keep model on GPU. soft_unload: Move to CPU, keep cached. full_unload: Remove from memory.
    enable_vae_tilingoptBOOLEANfalseEnable VAE tiling for large images. Reduces VRAM but slower.
    flow_shiftoptFLOAT2.80–10Flow-matching shift. Default 2.8 is balanced. Presets: portraits/faces 2.0–2.5 (sharper detail), landscapes/illustrations 3.5–5.0 (cleaner gradients, less high-frequency noise).
    reserve_vram_gboptFLOAT0.00–48Reserve VRAM for downstream nodes (upscalers, other models).
    moe_drop_tokensoptBOOLEANtrueTrue (default): MoE drops tokens that exceed expert capacity (lower VRAM, ~1–3% quality cost on dense regions). False: route every token through its top-K experts (best quality, higher VRAM peak — recommended only on ≥48GB cards).
    vae_dtypeoptCOMBObfloat16VAE decode precision. bfloat16 (default) is fast and matches model dtype. float32 reduces banding/chroma noise on smooth gradients with negligible cost on big cards. (Some users may already force this via ComfyUI launch flag.)
    force_reloadoptBOOLEANfalseForce full reload: clears cache, empties VRAM, reloads model fresh. Use if orphaned VRAM from failed loads.

    Outputs (2)

    NameTypeDescription
    imagesIMAGE
    final_promptSTRING