Open-H3-Regenerate-2K

An open replacement for the closed H3-Regenerate-2K API: a three-step MiniMax-H3 refinement pass that regenerates a 768P H3 draft (1344×768, 121 frames) at 2560×1440.

Project page · Code

Speed on one B200

One 5 s clip, 768P latent → 2560×1440 video: latent upscale + 3-step refine + VAE decode, warm, median of 3 runs:

Forward	VAE decoder	Total	Speedup
standard_fwd (dense attention, BF16)	released H3 VAE	87.26 s	1.00×
sol_engine_fwd (Sol-Attn, MXFP8, CUDA graphs)	released H3 VAE	37.39 s	2.33×
sol_engine_fwd (Sol-Attn, MXFP8, CUDA graphs)	fast H3 VAE decoder	22.45 s	3.89×

The fast H3 VAE decoder is the non-causal 24-layer decoder of StreamRender-H3, distilled from the released H3 VAE; it decodes the 2560×1440 clip in 1.6 s instead of 15 s. Latent upscale + refine alone go from 73.11 s to 20.46 s (3.57×).

Per-module numbers are in the code README.

Files

Path	Content
lora/open_h3_regenerate_2k_lora.safetensors	One merged LoRA (FP32, rank 576 on 312 linears, scale 1)
lora/sharp_curve_branches.safetensors	The 50 LMS Sharp curve AdaLN branches (FP32)
lora/lora_config.json, lora/module_ranks.json	Sources, hashes, per-module ranks
lora/MERGE_VERIFY.json	Merge check against the three separate adapters
latent_upsampler/	Author 3D-conv latent upscaler, BF16
vae_fast/fast_h3_vae_noncausal_24l.safetensors	Fast H3 VAE decoder from StreamRender-H3 (non-causal, 24 layers, 1.62B parameters)
base_model/	MiniMax-H3 Diffusers components (transformer, VAE, audio VAE, text encoder, tokenizer, processor, schedulers), unmodified copy of revision 6818f6c

base_model/ mirrors MiniMaxAI/MiniMax-H3 @ 6818f6c without changes (the native FL2VA/, Ref2VA/ and transformer_ref/ folders are left out). download.py in the code folder fetches everything here plus the optional HyperFlow 8-step adapter and Sharp's 1025×8 adaln_t_table (a base-derived tensor, read by byte range from Comfy-Org/MiniMax-H3).

What the LoRA contains

The reference recipe applies three adapters to the MiniMax-H3 transformer:

Adapter	Rank / alpha
Open-H3 refiner, step 5000 EMA	384 / 384
LightX2V FL2VA Turbo v1.2	128 / 8
LMS Sharp v1.0	64

They are merged by concatenating the factors along the rank axis with the strengths (and Turbo's alpha/rank) folded into B, so B @ A equals the sum of the three scaled deltas exactly. Over all 312 linears the merged delta matches the separate adapters to a relative error of 2.2e-8 (MERGE_VERIFY.json). Sharp's curve branches take an 8-D timestep curve instead of the layer input, so they are kept as separate modules.

Sampling: three Euler steps on sigmas [0.909375, 0.857143, 0.666667, 0], CFG 1, initial state (1 - 0.909375) * upscaled + 0.909375 * noise, model time 1 - sigma.

License

The refiner LoRA, the fast H3 VAE decoder and base_model/ follow the MiniMax H3 Community License, including its territory, use and redistribution terms. The components we did not train follow their own licenses: the latent upscaler, and the LightX2V FL2VA Turbo and LMS Sharp LoRAs merged into our LoRA. This bundle is not covered by a single blanket license. The runtime code's license does not replace model licenses.