Image Consistency — Same Person, Same Object, Same Style, Verified Locally

August 14, 2026

Generating a character is easy; keeping them the same across five images is the hard part. The new Image Consistency tool compares 2–4 session images in a single pass and returns a structured verdict — same person or not, same object or not, consistent visual style or drift — powered by Gemma 4 12B running locally on Ollama. Your images never leave the machine. The same engine now also audits the video editor's restyle pipeline automatically.

What you can do

• Compare 2–4 images in one call — pass filenames from your session's in/ or out/ folder and get a verdict with a confidence score, a one-line description of each image, and a list of specific differences ("Image 2's character has a different hair colour than Image 1").
• Pick the comparison subject — person (same person across photos?), object (same item, different angle?), style (same palette, rendering, lighting?), or auto to let the model infer what's common.
• Automatic restyle QA in the video editor — every restyle_frames / restyle_video run now samples up to 4 restyled frames, checks them for style and character drift, and attaches an advisory verdict to the result. It never blocks or re-runs your job; it tells you "possible palette drift between frames 3 and 7" so you can decide.
• A saved, searchable report — every comparison writes a markdown report to out/ and indexes it, so "what did the consistency check say" works in later turns.
• Fully local and free — runs on the production GPU server's Ollama, no external API, no per-call cost, and images stay on the box.

Where this shows up

• "I generated four shots of the same mascot — did Image Lab keep her consistent?"
• "Are these two product photos the same physical unit? Check the stickers and wear marks."
• "I restyled my clip as watercolour — did the look drift halfway through?"
• "Is the person in this ID photo the same person as in this event photo?"
• "Before I assemble the storyboard, verify the hero looks the same in all three keyframes."

Try it

• "Compare out/hero_v1.png and out/hero_v2.png — same character? Focus on hair, outfit, and face."
• "Check whether in/product_front.jpg and in/product_side.jpg show the same object."
• "Compare out/frame_001.png, out/frame_050.png, and out/frame_100.png for style drift."
• "Restyle my video as Studio Ghibli — and tell me what the consistency check says afterwards."

Heads up

• The verdict is advisory. Benchmarked on cross-image matching (August 2026), Gemma 4 12B is reliable but not infallible — treat borderline calls (confidence below ~70%) as a prompt to look yourself, not as ground truth.
• Max 4 images per call; order matters (the report refers to Image 1, Image 2, …).
• The first call in a while can take a few extra seconds while the model loads; subsequent calls run in ~2–5 seconds.