QuadForge / Wan 2.2 I2V / SVI 2.0 Pro / Prompt Enhancer / automated multi-part ComfyUI Workflow

v1.3:

• Prompt enhancer added that takes the input image as a reference (LLM-based, Gemma 3 abliterated).

  • Works for SFW as well as NSFW

  • Your own prompt can be very basic and short.

  • You can feed it back its output and write what you want to have changed, to - for example - include trigger words.

  • If needed, its custom instruction prompt can be adjusted in each "Enhancer" node.

  • If you dont want to use it, the group needs to be manuall bypassed for each part.

• Global LoRAs (SVI, lightx2v) only need to be loaded once.

• Aspect ratio of the input image is automatically taken over for the video.

  • You only need to adjust the desired resolution tier in the CustomResolution node.

Troubleshooting for the prompt enhancer if you get errors such as "LTXVGemmaClipModelLoader No package metadata was found for bitsandbytes":

• Update ComfyUI

• Update all custom nodes

• open cmd in your ComfyUI portable folder and type:

  • python_embedded\python.exe -m pip install -U bitsandbytes

  • or try: python_embedded\python.exe -m pip install -U bitsandbytes accelerate

Troubleshooting/best practice tips:

• The SVI-LoRA is mandatory in each part for the workflow to function (link in the workflow).

• The SVI-node does not seem to update with ComfyUI-manager, you`ll need to install it manually via Git (link in the workflow).

• For Sage Attention, this guide on YouTube helped me a lot: "How to Install Sage Attention 2.2 on Latest ComfyUI Portable And Deskop Version".

• Reduce the output bitrate to about 10 if you`re having trouble uploading to CivitAI.

• Start with 2x, 3x, 4x turned off and at a lower resolution (e.g. k20) and then go from there.

v1.2:

• Workflow reworked to support SVI 2.0 Pro, which uses the starting image and latent video from the previous clip to:

  • Improve transitions

  • Enable cross-clip memory for faces and other attributes

• Resolution can now only be set once for part 1; it is carried over to all subsequent clips

• Added switches for quickly switching between GGUF and diffusion/.safetensors models (e.g., SmoothMix). Deactivate LightX2v when using SmoothMix, as it is already embedded.

  • Note: SmoothMix changes faces and appearances quite a bit with this workflow compared to the GGUFs of base WAN.

• Sage Attention and patched Torch are now part of the model inputs and can be bypassed globally.

• Output folder structure updated: partial videos are now saved in the same folder as last-frame images.

• Separate switches for last frame and all frames image preview.

v1.1:

• Major declutter: Removed most crossing lines using set/get, subgraphs, and anything-everywhere nodes.

• Single model input propagated everywhere.

• Color-coded sections + cleaner layout.

CustomResolution Node Update:

• 9:16 & 16:9 support with shared resolution tiers.

• 24 & 30 FPS support.

• Auto-rounding to multiples of 8 for manual inputs.

https://github.com/AugustusLXIII/ComfyUI_CustomResolution_I2V/tree/main

v1.01:

Changes:

• Missing save last frame node added for scene 1.

• Switches for image preview (reduces generation time) and interpolation added.

Update: My CustomResolutionI2V node is now available via ComfyUI-manager. You might need to update ComfyUI-manager if you can`t find it.

QuadForge WAN 2.2 – ComfyUI Workflow

A powerful and flexible img2vid chaining workflow optimized for WAN 2.2 GGUF with the lightx2v-LoRA (6 steps).

Start with a single input reference image and generate up to four consecutive video segments. Each segment automatically uses the last frame of the previous clip as its starting image for perfect continuity. All segments are stitched together seamlessly, with final video interpolation (RIFE) applied as the last step for a smooth, high-quality output.

Key features:

• Adjustable number of generations (1–4) via simple switches

  • Tip: Start with 2x, 3x, and 4x turned off and add them until you like the result of the previous part. For this to work, you have to keep the seeds, LoRAs, and prompts fixed. Bypass the interpolation video node until you like your final result to save time.

• Independent prompt and LoRA control for each segment

• SageAttention enabled for approx. 40% speedup (bypass the two nodes in each step if not installed yet or if not supported by your GPU)

• Automatic last-frame extraction and saving after every generation

• Quick resolution and clip length adjustments using my own CustomResolution node (available via ComfyUI Manager): https://github.com/AugustusLXIII/ComfyUI_CustomResolution_I2V)

• Fully automated stitching

• RAM cleanup and model unload nodes for better stability

Ideal for creating long, consistent tracking shots, progressive scene reveals, or any multi-stage animated sequence from one starting image.

Notes:

• Lots of spaghetti but it works (will be polished in future updates)

• Framerate output in the CustomResolution node is currently fixed to 16 fps (I will make that adjustable as well)

• How-to manual: TBD