anime-forge

Lock art style and character design in keyframe images first, then use image-to-video only for motion like expressions and small actions. Keep style words out of i2v prompts so the model does not resample the look between frames.

Route tasks by tested behavior: GPT-family image generation handles frames containing Chinese text without garbling, while Gemini-family i2v follows N-second duration instructions in prompts. Subtitles and title cards render locally with Chrome and ffmpeg.

Google-family image and video models produce garbled Chinese characters, so generate any frame or card containing Chinese text with GPT-family image models instead. For subtitles, render text locally with Chrome and overlay it with ffmpeg for zero-cost, correct output.

Use anime-forge when the footage is generated rather than recorded, such as character-driven animated stories or IP short films. Screen recordings, tutorials, and product showcase videos belong to a recording-based workflow like video-forge.

Style descriptions in the i2v prompt invite resampling, causing jumps after the first frame. Derive style entirely from the starting frame, keep the prompt free of style words, and chain three to four timed actions to fill the clip.

Write a charter with scope and budget guardrails before generating anything, cap each prompt at three rolls, and build an animatic from still frames to validate pacing before paying for video generation. Static information cards skip i2v entirely.