LTX-2.3 Image-to-Video Workflow — QwenVL Auto-Prompt, No Drift — v1.0

This ComfyUI workflow turns still images into 24 FPS video, using QwenVL to draft motion prompts and saving video to ./output at 1920×1088.

QwenVL analyzes the input image and drafts a motion prompt with camera-lock and object-tracking hints; the listed camera setting is fixed, with no pan, zoom, or drift.

The workflow scales images to a 0.52 MP target, generates video at 960×544, then enlarges it 2× to 1920×1088.

Place images in ComfyUI’s ./input folder; you can review the generated motion prompt in QwenVL’s output text node before queuing.

The listing specifies 168 frames per generation at 24 FPS and says VideoHelperSuite saves the video in ./output.

The listing says model weights must be downloaded separately: six files totaling about 32 GB, placed in ComfyUI/models/unet, text_encoders, vae, and upscale_models.

Named model files include ltx-2.3-22b-distilled-Q4_K_M.gguf, gemma_3_12B_it_fp4_mixed.safetensors, LTX23_video_vae_bf16.safetensors, LTX23_audio_vae_bf16.safetensors, and ltx-2.3-spatial-upscaler-x2-1.1.safetensors.

The listing names these ComfyUI components: LTXV, AILab_QwenVL_Advanced, ComfyUI-GGUF, VideoHelperSuite, ImageIterator, ImageScaleToTotalPixels, and GetImageSize+; it marks rgthree-comfy optional.

Listed settings are er_sde sampling, 8 base steps, 3 refine steps, and CFG 1.0.

Tip: Edit QwenVL’s output text before queuing to guide motion direction, or bypass the Spatial Upscaler for 960×544 output.

Suggestion · not verified

The listing names ltx-2.3-spatial-upscaler-x2-1.1.safetensors, while the workflow file’s model list names ltx-2.3-spatial-upscaler-x2-1.0.safetensors; check which filename to use.

Suggestion · not verified

Before commercial use, review the LTX-2 Community License and Gemma Terms of Use; the listing notes an under-$10M annual-revenue condition for LTX-2 commercial use and an AI-content disclosure requirement.

Suggestion · not verified