LastEpisode/MiniMax-H3-FL-Singularity-Experimental

MiniMax H3 FL–Singularity Experimental

An experimental, reference-capable MiniMax H3 merge combining FL’s visual quality and clarity with Singularity’s more natural lighting and textures, expressive movement, and improved handling of distant faces. The aim is less waxiness than FL and less blur than standalone Singularity, while retaining useful strengths from both. Despite its FL foundation, this model supports reference-conditioned generation—not just first/last-frame workflows.

These are modified weights assembled from existing checkpoints, not a separately trained model.

What changed

Attention and MLPs: all 50 main transformer blocks use a 50/50 blend of FL and Singularity for these four projection weights:

blocks.<0..49>.attn.qkv_proj.weight
blocks.<0..49>.attn.out_proj.weight
blocks.<0..49>.mlp.fc1.weight
blocks.<0..49>.mlp.fc2.weight

The blend is FL + 0.5 × (Singularity − FL), calculated in FP32 and rounded to BF16 before quantization. Native tensor ordering is retained; Singularity's leading model.diffusion_model. key prefix is removed for correspondence.

Time/AdaLN: Singularity supplies its complete matching pruned bundle: adaln_t_table, all 50 blocks' adaln_proj.linear.weight and .bias, and the final layer's corresponding weight/bias pair. These 103 tensors are copied together, preserving their values and dtypes, rather than mixing different timestep coordinate systems.

Everything else remains FL: the conditioning projection, both token-refiner blocks, normalization gains, video/audio input and output projections, and rotary frequencies. This is not a whole-model 50/50 blend.

Sources

From Comfy-Org/MiniMax-H3 at the build revision:

• minimax_h3_fl2va_pruned_bf16.safetensors — floating FL weights for the blend.
• minimax_h3_fl2va_pruned_int8_convrot.safetensors — deployment base; untouched tensors are retained byte-for-byte.

From WarmBloodAban/Minimax-h3_Singularity at the build revision:

• Minimax-h3_Singularity_ref2va_v2_bf16.safetensors — floating attention/MLP donor.
• Minimax-h3_Singularity_ref2va_v2_pruned_int8.safetensors — matching timestep/AdaLN donor.

Weight construction

Only the 200 blended main-block projection matrices are converted to INT8 ConvRot, replacing their weights and per-row scales in the deployed FL base. The conversion uses Comfy-Org's quantizer at this revision: group size 256, per-row absmax scaling, CPU FP32 arithmetic, 256 complete rows per chunk, and mseclip=False.

The existing quantization configuration is retained. The token refiner is not quantized or replaced. Singularity's complete time/AdaLN bundle is then copied into the result. No additional pruning is performed.

Credits and license

Built from work by MiniMax, Comfy-Org, and WarmBloodAban / Singularity. The modified checkpoint remains subject to the MiniMax H3 Community License Agreement; see LICENSE and NOTICE.