MiniMax H3-Base Ref2VA field study: input roles, character identity, shot timing, and measured A/Bs
I run the released MiniMax H3-Base Ref2VA on rented GPUs for original-character video production. I expanded these field notes into a technical study with accepted and rejected take IDs, our Ref2VA input-role map, Qwen3-VL encoder probes, shot-timing comparisons, measured render costs, failure analysis, and two side-by-side MP4s. The article and evidence ledger are linked below.
Scope and terminology. FL2VA and Ref2VA are the released H3-Base checkpoints. “Character swap” and “world swap” are task routes in my Studio, not official H3 variants. Our local comparisons did not use hosted H3-Context-IR or H3-Regenerate-2K.
How the inputs are assembled. In our local Ref2VA path, numbered visual references are assembled before the written task for the causal Qwen encoder. Image and video content also reaches H3 through VAE latents; the later task text still matters to the generator. We measured image-encoding similarity across later prompts and used a saved-input replay to separate encoder work from H3 generation. This is a finding about our pipeline and probes, with the implementation details and limits documented in the study.
Reference-video roles and identity. When the source video was used as a visual reference, it helped preserve motion and location for a performer replacement, but pulled a world-change attempt back toward the source room. For one accepted six-second first segment of a world change, we instead used the source for motion/camera guidance and supplied target-world guide frames at shot cuts. The rest of that film remains unaccepted. In character work, a face-only input left the source body and costume; a full-character input produced an accepted ten-second single-character replacement. A two-character take improved with coherent references plus an explicit shot plan. Both the reference package and the plan changed, so this is a workflow comparison, not an isolated image-count result.
Timing and caching. In one same-source timing repair, writing a cut as “At 00:SS.mmm, the shot cuts to…” improved alignment. Ranges worked in other takes, so the direct timing comparison (n=2) does not establish a universal prompt rule. A separate same-seed, same-input, 24-step Safe-cache A/B on one accepted scene reduced total render time from 934.5 to 717.8 seconds, about 23%. That is one scene and setup, not a general speed guarantee. The side-by-side MP4s let readers inspect these comparisons directly.
Field study and MP4 links: https://github.com/Klimentios/Klimentios/blob/main/writing/minimax-h3-field-notes.md
Evidence ledger and measurement boundaries: https://github.com/Klimentios/Klimentios/blob/main/writing/minimax-h3-evidence.md
The source footage and detailed Studio job logs are private. The public article gives the method, selected measurements, take IDs, and visual comparisons; it is not a fully reproducible public dataset. I would welcome comparisons from other local Ref2VA users, especially on input roles and source-aligned cuts.