### H3 ErosLite — 8-Step Turbo (NVFP4 Smart-Mixed Edition)
H3 ErosLite is a high-performance, edge-preserved 8-step Turbo video generation checkpoint built for the MiniMax H3 architecture.
It is specifically engineered to fit within 16GB consumer GPUs (such as RTX 4080/4090 and the new RTX 50-series Blackwell cards) while completely eliminating the quality loss, blurriness, and washed-out "matte" look typical of standard 4-bit quantizations.
---
#### 🌟 What is "Smart-Mixed" Precision?
Standard 4-bit quantizations compress all transformer layers equally, which severely damages microcontrast, softens facial textures, and blurs hair/skin details.
H3 ErosLite implements Lossless Edge Preservation:
1. Blocks 0 & 1 (Input / Early Feature Extraction): Preserved in 100% Lossless BF16 to ensure zero noise at initial edge, eye, and contour processing.
2. Blocks 47, 48 & 49 (Output / Latent Reconstruction): Preserved in 100% Lossless BF16 to synthesize crisp skin pores, high-frequency details, and cinematic dynamic range.
3. Blocks 2 through 46 (Intermediate Layers — 180 Linears): Quantized to native NVFP4 (FP4-E2M1 with FP8 block scales) to leverage GPU Tensor Cores for ultra-fast inference and minimal VRAM consumption.
4. All Embeddings, Norms, Biases & Patch Projections: Preserved in native BF16/FP32.
The result is a compact ~14.2 GB checkpoint that delivers crisp, cinema-grade sharpness and deep contrast indistinguishable from full 40GB uncompressed weights.
---
#### ⚙️ Recommended Settings (ComfyUI)
- Model Loader: Standard UNETLoader or Load Diffusion Model (loads natively with ComfyUI mixed-precision ops).
- Steps: 8 steps (distilled turbo architecture; do not exceed 10).
- Sampler: euler or euler_ancestral
- Scheduler: simple or sgm_uniform
- CFG / Guidance: 1.0 to 1.2 (do NOT use CFG > 1.5 on turbo models).
- Video Shift: 12 (critical for smooth motion and flicker prevention).
- Audio Shift: 3 (for native synchronized lip-sync and audio).
- Denoise (2-Pass Upscaler): Keep between 0.18 and 0.22 if using a secondary latent pass to preserve facial identity.
---
#### 🎬 Prompting Structure (MiniMax H3 Format)
H3 ErosLite excels in both Text-to-Video (T2V) and Image-to-Video (I2V) workflows. For multi-shot scenes with dialogue, use the standard MiniMax H3 multimodal format:
```text
integrated_multimodal_description:
[Shot 1] Cinematic medium close-up, 4k ultra-detailed. A captivating scene featuring... The camera performs a smooth slow push-in as she smiles seductively at the lens. At 00:02.000, she parts her lips and says: <d>[English] "Your dialogue here..."</d>
[Shot 2] At 00:03.200, smooth cut to a close-up tracking shot. The camera orbits gently while maintaining natural eye contact until the end of the scene.
overall_soundscape:
Intimate room tone, soft breathing, and faint ambient sound.
non_diegetic_music:
A slow, atmospheric ambient synth layered with gentle acoustic tones.
```
---
Compatible with ComfyUI, diffusers pipelines, and LoRAs trained for MiniMax H3.