SocksLatentPatcher GGUF Subgraph — v2.0

This workflow is for experimental image-to-video generation with Wan Video 2.2 I2V-A14B; it uses `wan2.2_i2v_high_noise_14B_Q4_K_S.gguf` and `wan2.2_i2v_low_noise_14B_Q4_K_S.gguf`, patches latent tensors instead of decoding through the VAE, and can produce video with interpolated

The subgraph combines PainterI2V, SocksLatentPatcher, Rife Tensorrt, and Upscaler Tensorrt.

It bypasses VAE decode and directly patches the latent tensor; the workflow description presents this as a way to handle detail loss and saturation or desaturation during infinite-length video generation.

The main model inputs are `wan2.2_i2v_high_noise_14B_Q4_K_S.gguf` and `wan2.2_i2v_low_noise_14B_Q4_K_S.gguf`.

The video output path includes interpolation and upscaling options.

Links for model, CLIP, and VAE files are included in the workflow.

This is version v2.0, published by Donzo15.

Listed model files include `wan2.2_i2v_high_noise_14B_Q4_K_S.gguf`, `wan2.2_i2v_low_noise_14B_Q4_K_S.gguf`, `lightx2v_I2V_14B_480p_cfg_step_distill_rank256_bf16.safetensors`, and `Wan_2_2_I2V_A14B_HIGH_lightx2v_4step_lora_v1030_rank_64_bf16.safetensors`.

Other listed model files are `nsfw_wan_umt5-xxl_fp8_scaled.safetensors`, `wan_2.1_vae.safetensors`, `rife49.pth`, and `film_net_fp32.pt`.

Listed node packs include `comfyui_memory_cleanup`, `comfyui-videohelpersuite`, `rgthree-comfy`, `ComfyUI-GGUF`, `ComfyUI-Rife-Tensorrt`, `ComfyUI-Upscaler-Tensorrt`, `comfyui-frame-interpolation`, `comfyui_essentials`, `comfyui-kjnodes`, and `comfyui-easy-use`.

For interpolation, choose between `rife vf1` and `rife tensorrt`.

The I2V path is marked experimental.

Before running, verify the exact `ComfyUI-GGUF` package and confirm that these files are available: `lightx2v_I2V_14B_480p_cfg_step_distill_rank256_bf16.safetensors`, `Wan_2_2_I2V_A14B_HIGH_lightx2v_4step_lora_v1030_rank_64_bf16.safetensors`, `nsfw_wan_umt5-xxl_fp8_scaled.safet v

Vorschlag · nicht geprüft