LTX Flow - Qwen Edit Bridge

The name sounds like it should be calling some hosted API. It isn't. Qwen Edit Bridge is pure local glue: it takes two images - your original frame and a Qwen Image Edit version of it - and turns them into matching first/last keyframes for LTX 2.3. No network, no key, nothing to sign up for.

Here's the workflow it lives in, because that's the only way the name makes sense. Qwen-Image Edit 2511 rewrites an image locally: change the sofa's fabric, swap the car's color, add an object to the scene. You now have "before" and "after" stills. The Bridge sends the before as LTX's first frame and the after as the last frame, and LTX generates the motion that turns one into the other. That's the pack's headline "Qwen Image Edit to LTX" pipeline, and it's genuinely fun - the model does the transformation, LTX does the animation, and you never touch a remote server.

How it works

For each input it keeps just the first frame and the first three channels (RGB), because these are stills, not clips. Then it figures out a target size - resize_to picks which image sets the dimensions - and floors width and height down to a multiple of multiple_of (32 by default). That multiple isn't cosmetic: LTX 2.3 requires dimensions divisible by 32, and off-grid values fail rather than round. Both images get resized with lanczos and center-cropped to match. first_frame_source decides which one opens the clip; the edit is always the last frame. Output keyframe_batch is just the two frames concatenated into one IMAGE batch, handy if you want to eyeball both at once.

The inputs that matter

• source_image / qwen_edited_image - your original and the Qwen-edited version.
• first_frame_source - which image starts the clip. Default is the source, so LTX animates into the edit. Flip it if you want the reverse.
• resize_to - which image's dimensions win. Defaults to the source.
• multiple_of - 32. Leave it. Set to 0 to disable rounding, but then make sure your dims are already 2.3-legal.

Outputs: first_frame, last_frame, keyframe_batch - wire the first two into First/Last Guide's slots.

Install and what it actually costs you

Search "comfyui-extend" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/balarooty/comfyui-extend

Restart ComfyUI. The Bridge node itself has zero dependencies - it's resize + concat. What makes this workflow heavy is everything around it: the Qwen-Image Edit diffusion model, the qwen_2.5_vl_7b text encoder, the Qwen VAE, then the LTX 2.3 checkpoint and Gemma encoder on top. Plus the workflow's other custom packs: ComfyUI-KJNodes, WhatDreamsCost-ComfyUI, and ComfyUI Easy Use. The repo's scripts/download_models.sh grabs the model files (with COMFYUI_DIR pointed at your ComfyUI folder) and scripts/install_custom_nodes.sh clones the three packs.

Common issues

• Version mismatch between README and script. The README and the workflow's own notes reference Qwen-Image 2511 files (qwen_image_edit_2511_bf16.safetensors, Qwen-Image-Edit-2511-Lightning-4steps), but download_models.sh pulls the 2509 variants (qwen_image_edit_2509_fp8_e4m3fn.safetensors, Qwen-Image-Edit-2509-Lightning). If the workflow complains about a missing model after you ran the script, that's why - grab the 2511 files or rename what you have.
• "Invalid resolution." Your source images aren't landing on a multiple of 32 after resize. Keep multiple_of at 32 and check the source dimensions.
• It's for stills, not video. Only the first frame of each input is used. Feeding it clips wastes everything past frame one.

workflows/05_qwen_edit_to_ltx_first_last.json is the full pipeline. It's the most demanding workflow in the pack, so don't make it your first stop - get a plain first/last clip working before you stack Qwen on top.