Portuguese (pt-BR) speech degrades vs. MiniMax-H3 in ref2va — which languages were used for pruning/distillation?
Hi, thanks for releasing HyperFlow H3 Flash.
We ran a small paired test of ref2va on a Brazilian Portuguese talking-head scene and saw a consistent drop in speech articulation.
Setup
- ComfyUI v0.42.3, PyTorch 2.11 (cu130), RTX 5090.
- HyperFlow:
minimax_h3_ref2va_hyperflow_flash_int8_convrot.safetensors(BennyDaBall's INT8 port), HyperFlow INT8 Loader (legacy patcher) + ComfyUI-HyperFlow-Flash 1.1.1, HyperFlow Sigmas 6 steps, euler, BasicGuider (no CFG). The README ref2va example (seed 42) reproduces almost frame-for-frame on our machine, so the install looks correct. - Baseline: a MiniMax-H3 ref2va derivative (Singularity v2 pruned INT8) + a turbo LoRA, 6 steps.
- Everything else identical: same prompt (spoken lines as
<d>[Portuguese] …</d>), same reference images, same 4 s voice reference, same reference video, first-frame guide, 704×1248, 243 frames, same seeds.
Result (3 seeds per scene, faster-whisper large-v3 on the isolated audio)
- 10 s monologue (32 words): HyperFlow broke words in 3/3 seeds, baseline in 0/3. "original" → "Orgal" / "Origal" / "Orgnal" (3/3), "estampa" → "estafa" / "estapas", "flanela" → "flayela".
- 8 s close-up (18 words): 1/3 vs 0/3.
- The video side was stable: subject stays in frame, identity and props fine.
Question: the MiniMax-H3 card lists Portuguese among its 11 stable dialogue languages, while this card is tagged en/zh. Which languages or data were used for the pruning calibration and the two-time distillation? Is non-EN/ZH speech expected to degrade? Would 8 steps or the official INT8 package behave differently? Happy to share more details.
spanish is broken too
Hi @MagnusVector and @alexloops — thanks for the detailed paired benchmark and for sharing your setup.
What you are seeing is an expected capacity trade-off of the 14B Flash compression. Cutting the ~31B transformer down by more than half (to 13.9B via block pruning and low-rank factorization) preserves visual identity and motion well, while fine-grained speech articulation in languages like Portuguese and Spanish is naturally more sensitive to parameter reduction.
Two recommendations depending on your workflow:
- Try 8 steps on Flash (
HyperFlow Sigmas = 8): The 8-step schedule is our quality preset and gives the audio branch extra denoising refinement over the 6-step default. (Note: our officialref2va_int8_convrot_comfyui/weights share the same INT8 codes and scales as the port you tested, so switching the file alone won't change articulation.) - Use full-size
VideoRebirthAI/hyperflow(GitHub):Flash (14B)is built specifically for maximum speed and low VRAM. For dialogue-heavy scenes where full 31B articulation fidelity is the top priority, we recommend our 8-stepHyperFlowLoRA on the unpruned 31B MiniMax-H3 base.