VRGDG LTX IC Ingredients Grid

VRGDG_LTXICIngredientsGrid composes up to 24 images into a single reference sheet built for LTX InContext LoRA conditioning. If you're using LTX's "Ingredients" IC-LoRA - the one that conditions generation on a reference sheet so characters, props, and locations carry across clips - this is the node that builds that sheet.

To get why this exists, you need the wider LTX picture. Character consistency across shots is the model family's sore spot: faces and costumes drift from clip to clip, and the community's fix has been moving from training toward conditioning on references. Lightricks' Ingredients IC-LoRA is exactly that - you give it a sheet of reference images (your character, her outfits, key props, the locations) and it conditions every generated clip on that sheet, keeping the look stable across scenes. That sheet needs to be one carefully arranged image at a specific aspect ratio, which is a fiddly thing to assemble by hand. This node does it for you, and does it dynamically from whatever images are upstream in your graph.

Inputs that matter

• image1 … image24 - the ingredient images: character stills, outfit shots, prop close-ups, location frames. The tooltips call each one an "Ingredient image," which tells you exactly what slot it fills.
• image_count - how many of those 24 sockets to show and compose (default 6). You rarely need all 24; 4–8 ingredients is the sweet spot for a character + outfits + prop sheet.
• layout - the composition preset. auto_ltx is the default and picks a sensible arrangement; the others (six_panel_story, three_row_reference, uniform_grid, horizontal_strip, …) are preset grids for different sheet styles. uniform_grid with columns set is the one to reach for when you want total control.
• output_width / output_height - the sheet resolution, default 768×448, which is LTX's native aspect. Feed the IC-LoRA its preferred resolution and it behaves noticeably better.
• fit_mode - contain_pad (letterbox, keeps whole image) vs cover_crop (fills the cell, crops edges). Default contain_pad; swap to cover_crop if you want no letterbox bars.
• gutter, outer_padding, corner_radius - spacing and rounded-corner polish between cells.
• batch_mode - first_image_only (build one sheet from the first frame of each input batch) vs all_images (build a sheet per frame set). Default first_image_only keeps VRAM sane for video inputs.

Output is the single reference_sheet image, which wires into the Ingredients IC-LoRA's conditioning/control input.

Why it's worth having

Making one of these sheets manually is a sub-project: aligning images, matching aspect, saving it, reloading it. Doing it as a node means the sheet rebuilds whenever your upstream images change - swap an outfit still and the whole sheet updates before the next run. In a music video, where the same character needs to look right across dozens of scenes, this is the difference between a coherent video and a cast of unrelated lookalikes.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl

Restart ComfyUI; it's under VRGDG/LTX. ComfyUI Manager: search vrgamedev. Image composition here leans on kornia (one of the three packages in requirements.txt), so run pip install -r requirements.txt from the repo folder if imports fail.

Honest note

Two things to know. The sheet only helps if the downstream IC-LoRA setup expects this exact style of reference grid - if your LTX version's Ingredients conditioning wants a different sheet format, match its docs rather than assuming. And more images isn't automatically better: throw 24 cluttered ingredients at it and the IC-LoRA has no idea what to focus on. A clean 4–8 image sheet with one clear subject per cell beats a crowded collage every time.