# 🎬 H3 ErosLite: Minimalist 2-Pass Video Workflow & Technical Guide
> **Author:** 4lth3rego
> **Base Model:** MiniMax H3 (ErosLite NVFP4 Turbo)
> **Target Framework:** ComfyUI
> **Pipeline:** 2-Pass Hybrid Generation (Base Turbo + 3D Latent Upscale + Cinema Sharpen)
---
## 🌟 1. Overview & Architectural Philosophy
Generating high-definition AI video on consumer hardware traditionally faces two severe bottlenecks:
1. **Extreme VRAM Consumption:** High-resolution video diffusion natively requires massive memory buffers, often causing Out-Of-Memory (OOM) errors on 16 GB GPUs.
2. **Facial Distortion in Upscaling:** Standard spatial or latent upscalers frequently hallucinate micro-textures, altering facial identity and introducing temporal flickering between frames.
The **H3 ErosLite 2-Pass Workflow** solves both challenges through an engineered, modular pipeline:
* **Stage 1 (Base Motion Generation):** Generates fluid, temporally coherent video at **0.4 MP (~864×480)** using **8 Turbo steps** at lightning speed.
* **Stage 2 (3D Latent Upscaling):** Uses MMH3UltimateUpscale with a 3D latent diffusion model and a 1-step solver at a strictly calibrated denoise of **0.18**, upscaling to **0.9 – 1.0 MP (~1344×768)** with zero facial morphing.
* **Stage 3 (Cinema Sharpening):** Applies a sub-second edge-refinement pass that brings out cinematic film clarity without artificial grain.
* **Canvas Modularization:** The entire pipeline is organized into **5 clean Subgraphs**, reducing 60+ cluttered nodes into an intuitive, elegant workspace.
---
## 🧩 2. Detailed Breakdown of Workflow Subgraphs
### 📦 Subgraph 1: H3 ErosLite Model Loader
* **Primary Role:** Manages model weight ingestion and VRAM offloading.
* **Checkpoint:** Loads H3_ErosLite.safetensors (14.25 GiB). Edge layers (blocks 0, 1, 47, 48, 49) are preserved in lossless **BF16**, while intermediate linear layers are quantized to **NVFP4 (FP4-E2M1 with FP8 block scales)**.
* **Text & Conditioning:** Loads qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors via CLIPLoader.
* **VAEs:** Separately loads the Video VAE minimax_h3_video_vae_fp16.safetensors) and Audio VAE minimax_h3_audio_vae_fp32.safetensors).
* **VRAM Footprint:** Fits comfortably on 16 GB GPUs (RTX 4080, RTX 5060 Ti, RTX 4070 Ti, RTX 3090).
---
### ⏱️ Subgraph 2: Duration to Frame Length
* **Primary Role:** Enforces mathematical compatibility with MiniMax H3's temporal compression.
* **The 17n + 5 Rule:** MiniMax H3's 3D VAE requires video frames to strictly conform to the formula:
$$\text{Frames} = \max(5, \text{round}(t \times 24)) + \left(5 - (\max(5, \text{round}(t \times 24)) \pmod{17})\right) \pmod{17}$$
* **Usage:** Enter duration in seconds (e.g. 5.0s). The subgraph automatically computes **124 frames** (at 24 fps), preventing VAE slicing errors.
---
### 📐 Subgraph 3: Resolution Standardizer
* **Primary Role:** Standardizes input geometry across different aspect ratios and generation passes.
* **Core Functions:**
* **Aspect Ratio Selection:** Select standard presets 16:9 widescreen 1344x768, 9:16 vertical, 1:1 square, 5:8 portrait, etc.) or pass original input dimensions.
* **Pass 1 Sizing (0.4 MP):** Scales input to ~400,000 pixels (e.g., 864×480) strictly rounded to multiples of 32 for the base motion pass.
* **Pass 2 Sizing (0.9 – 1.0 MP):** Computes the proportional upscale target (e.g., 1344×768) and directly drives the 3D upscaler parameters.
* **Mode Switch:** Toggle 1 for Original Image Aspect Ratio, or 2 for Aspect Ratio Override / Resize.
---
### ✍️ Subgraph 4: H3 Promptor (Multimodal Engine)
* **Primary Role:** Translates and structures creative concepts into MiniMax H3's native multimodal prompt syntax.
* **Internal Stages:**
1. **🌐 Translation:** GoogleTranslateTextNode translates non-English inputs to clean English.
2. **👁️ Vision Feature Extraction:** H3_Vision_Analyzer analyzes the initial frame image image_0) to extract character identity, clothing, lighting, and palette.
3. **🧹 VRAM Purge:** Triggers LayerUtility: PurgeVRAM V2 to free memory before diffusion starts.
4. **✍️ Multimodal Structuring:** H3_Promptor synthesizes the vision context and creative description into official MiniMax tags:
* <d>[Language] ... </d>
* integrated_multimodal_description:
* overall_soundscape:
* non_diegetic_music:
5. **👁️ Live Preview:** The generated prompt is routed to a ShowText node for immediate inspection.
---
### 🔍 Subgraph 5: MMH3 2-Pass Latent 3D Upscaler
* **Primary Role:** High-resolution spatial and micro-texture refinement.
* **Engine:** MMH3UltimateUpscale paired with minimax_h3_latent_upscaler_3d_bf16.safetensors.
* **Identity Preservation Setting:** Calibrated to **1 step** with beta57 scheduler at **denoise = 0.18**. This guarantees that skin pores, hair strands, and lighting reflections are crisply added without altering facial anatomy or character identity.
* **Cinema Sharpening:** Output frames pass through ImageSharpen (Radius: 1, Sigma: 0.8, Strength: 0.35) to remove digital softness and deliver filmic texture.
---
## 🚀 3. Step-by-Step User Guide
### Mode A: Image-to-Video (I2V) — Recommended
1. Drag and drop your source image into *1. Start Frame**.
2. Set your desired duration in *⏱️ Duration** (default: 5.0s).
3. In *✍️ H3 Promptor**, type a description of the desired action (e.g. "A woman smiling and turning her head towards the camera on a windy beach").
4. Leave *2. End Frame** empty.
5. Click **Queue Prompt**.
---
### Mode B: First & Last Frame (FL2V)
1. Load your beginning image into *1. Start Frame**.
2. Load your target finishing image into *2. End Frame**.
3. In *✍️ H3 Promptor**, set Task Type to FL2VA or describe the transitional trajectory.
4. Click **Queue Prompt**. The model will interpolate fluid motion connecting both keyframes.
---
### Mode C: Reference-to-Video (Ref2V)
1. Load your reference character/style image into *3. Reference Image**.
2. In **Group 4**, disconnect the latent from MiniMaxH3ImageToVideo and connect the latent output from MiniMaxH3ReferenceToVideo (Node 10) to the Sampler.
3. Click **Queue Prompt**.
---
## ⚙️ 4. Recommended Parameter Reference Table
| Parameter | Pass 1 (Base Generation) | Pass 2 (Latent 3D Upscale) | Notes |
| :--- | :--- | :--- | :--- |
| **Resolution** | ~864×480 (0.4 MP) | ~1344×768 (0.9 – 1.0 MP) | Driven automatically by Subgraph 3 |
| **Sampling Steps** | 8 Turbo Steps | 1 Step | Fast 8-step base + 1-step upscale |
| **Sampler** | euler | er_sde or sa_solver | Ultra-stable temporal coherence |
| **Scheduler** | beta57 | beta57 | Calibrated MiniMax curve |
| **Denoise** | 1.00 | *0.18 – 0.20** | **Do not exceed 0.25** to avoid face morphing |
| **Frame Rate** | 24 fps | 24 fps | Standard cinema frame rate |
| **Sharpening** | None | Radius: 1, Sigma: 0.8, Strength: 0.35 | Sub-second filmic enhancement |
---
## 💡 5. Best Practices & Pro Tips
* **Preserving Facial Likeness:** Keep Pass 2 denoise strictly between 0.16 and 0.20. A denoise of 0.50 will rewrite facial features and destroy identity.
* **Audio Synchronization:** The Audio VAE synthesizes a synchronized soundscape directly from the multimodal prompt. Use headphones to preview ambient audio and spatial sound effects.
* **Prompting Secret:** Let H3 Promptor do the heavy lifting. Avoid pasting giant walls of raw tags; instead, describe natural camera movement (*"slow tracking shot"*, "medium close-up") and physical dynamics (*"hair blowing gently in the wind"*).
---
## 📥 6. Installation & Files
1. Place the model checkpoint in:
ComfyUI/models/diffusion_models/H3_ErosLite.safetensors
2. Place the CLIP model in:
ComfyUI/models/clip/qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors
3. Place the VAEs in:
ComfyUI/models/vae/minimax_h3_video_vae_fp16.safetensors
ComfyUI/models/vae/minimax_h3_audio_vae_fp32.safetensors
4. Drag and drop the workflow file H3_ErosLite_2Pass_Clean_Workflow.json into ComfyUI.
Description
# 🎬 H3 ErosLite: Minimalist 2-Pass Video Workflow & Technical Guide
> **Author:** 4lth3rego
> **Base Model:** MiniMax H3 (ErosLite NVFP4 Turbo)
> **Target Framework:** ComfyUI
> **Pipeline:** 2-Pass Hybrid Generation (Base Turbo + 3D Latent Upscale + Cinema Sharpen)
---
## 🌟 1. Overview & Architectural Philosophy
Generating high-definition AI video on consumer hardware traditionally faces two severe bottlenecks:
1. **Extreme VRAM Consumption:** High-resolution video diffusion natively requires massive memory buffers, often causing Out-Of-Memory (OOM) errors on 16 GB GPUs.
2. **Facial Distortion in Upscaling:** Standard spatial or latent upscalers frequently hallucinate micro-textures, altering facial identity and introducing temporal flickering between frames.
The **H3 ErosLite 2-Pass Workflow** solves both challenges through an engineered, modular pipeline:
* **Stage 1 (Base Motion Generation):** Generates fluid, temporally coherent video at **0.4 MP (~864×480)** using **8 Turbo steps** at lightning speed.
* **Stage 2 (3D Latent Upscaling):** Uses MMH3UltimateUpscale with a 3D latent diffusion model and a 1-step solver at a strictly calibrated denoise of **0.18**, upscaling to **0.9 – 1.0 MP (~1344×768)** with zero facial morphing.
* **Stage 3 (Cinema Sharpening):** Applies a sub-second edge-refinement pass that brings out cinematic film clarity without artificial grain.
* **Canvas Modularization:** The entire pipeline is organized into **5 clean Subgraphs**, reducing 60+ cluttered nodes into an intuitive, elegant workspace.
---
## 🧩 2. Detailed Breakdown of Workflow Subgraphs
### 📦 Subgraph 1: H3 ErosLite Model Loader
* **Primary Role:** Manages model weight ingestion and VRAM offloading.
* **Checkpoint:** Loads H3_ErosLite.safetensors (14.25 GiB). Edge layers (blocks 0, 1, 47, 48, 49) are preserved in lossless **BF16**, while intermediate linear layers are quantized to **NVFP4 (FP4-E2M1 with FP8 block scales)**.
* **Text & Conditioning:** Loads qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors via CLIPLoader.
* **VAEs:** Separately loads the Video VAE minimax_h3_video_vae_fp16.safetensors) and Audio VAE minimax_h3_audio_vae_fp32.safetensors).
* **VRAM Footprint:** Fits comfortably on 16 GB GPUs (RTX 4080, RTX 5060 Ti, RTX 4070 Ti, RTX 3090).
---
### ⏱️ Subgraph 2: Duration to Frame Length
* **Primary Role:** Enforces mathematical compatibility with MiniMax H3's temporal compression.
* **The 17n + 5 Rule:** MiniMax H3's 3D VAE requires video frames to strictly conform to the formula:
$$\text{Frames} = \max(5, \text{round}(t \times 24)) + \left(5 - (\max(5, \text{round}(t \times 24)) \pmod{17})\right) \pmod{17}$$
* **Usage:** Enter duration in seconds (e.g. 5.0s). The subgraph automatically computes **124 frames** (at 24 fps), preventing VAE slicing errors.
---
### 📐 Subgraph 3: Resolution Standardizer
* **Primary Role:** Standardizes input geometry across different aspect ratios and generation passes.
* **Core Functions:**
* **Aspect Ratio Selection:** Select standard presets 16:9 widescreen 1344x768, 9:16 vertical, 1:1 square, 5:8 portrait, etc.) or pass original input dimensions.
* **Pass 1 Sizing (0.4 MP):** Scales input to ~400,000 pixels (e.g., 864×480) strictly rounded to multiples of 32 for the base motion pass.
* **Pass 2 Sizing (0.9 – 1.0 MP):** Computes the proportional upscale target (e.g., 1344×768) and directly drives the 3D upscaler parameters.
* **Mode Switch:** Toggle 1 for Original Image Aspect Ratio, or 2 for Aspect Ratio Override / Resize.
---
### ✍️ Subgraph 4: H3 Promptor (Multimodal Engine)
* **Primary Role:** Translates and structures creative concepts into MiniMax H3's native multimodal prompt syntax.
* **Internal Stages:**
1. **🌐 Translation:** GoogleTranslateTextNode translates non-English inputs to clean English.
2. **👁️ Vision Feature Extraction:** H3_Vision_Analyzer analyzes the initial frame image image_0) to extract character identity, clothing, lighting, and palette.
3. **🧹 VRAM Purge:** Triggers LayerUtility: PurgeVRAM V2 to free memory before diffusion starts.
4. **✍️ Multimodal Structuring:** H3_Promptor synthesizes the vision context and creative description into official MiniMax tags:
* <d>[Language] ... </d>
* integrated_multimodal_description:
* overall_soundscape:
* non_diegetic_music:
5. **👁️ Live Preview:** The generated prompt is routed to a ShowText node for immediate inspection.
---
### 🔍 Subgraph 5: MMH3 2-Pass Latent 3D Upscaler
* **Primary Role:** High-resolution spatial and micro-texture refinement.
* **Engine:** MMH3UltimateUpscale paired with minimax_h3_latent_upscaler_3d_bf16.safetensors.
* **Identity Preservation Setting:** Calibrated to **1 step** with beta57 scheduler at **denoise = 0.18**. This guarantees that skin pores, hair strands, and lighting reflections are crisply added without altering facial anatomy or character identity.
* **Cinema Sharpening:** Output frames pass through ImageSharpen (Radius: 1, Sigma: 0.8, Strength: 0.35) to remove digital softness and deliver filmic texture.
---
## 🚀 3. Step-by-Step User Guide
### Mode A: Image-to-Video (I2V) — Recommended
1. Drag and drop your source image into *1. Start Frame**.
2. Set your desired duration in *⏱️ Duration** (default: 5.0s).
3. In *✍️ H3 Promptor**, type a description of the desired action (e.g. "A woman smiling and turning her head towards the camera on a windy beach").
4. Leave *2. End Frame** empty.
5. Click **Queue Prompt**.
---
### Mode B: First & Last Frame (FL2V)
1. Load your beginning image into *1. Start Frame**.
2. Load your target finishing image into *2. End Frame**.
3. In *✍️ H3 Promptor**, set Task Type to FL2VA or describe the transitional trajectory.
4. Click **Queue Prompt**. The model will interpolate fluid motion connecting both keyframes.
---
### Mode C: Reference-to-Video (Ref2V)
1. Load your reference character/style image into *3. Reference Image**.
2. In **Group 4**, disconnect the latent from MiniMaxH3ImageToVideo and connect the latent output from MiniMaxH3ReferenceToVideo (Node 10) to the Sampler.
3. Click **Queue Prompt**.
---
## ⚙️ 4. Recommended Parameter Reference Table
| Parameter | Pass 1 (Base Generation) | Pass 2 (Latent 3D Upscale) | Notes |
| :--- | :--- | :--- | :--- |
| **Resolution** | ~864×480 (0.4 MP) | ~1344×768 (0.9 – 1.0 MP) | Driven automatically by Subgraph 3 |
| **Sampling Steps** | 8 Turbo Steps | 1 Step | Fast 8-step base + 1-step upscale |
| **Sampler** | euler | er_sde or sa_solver | Ultra-stable temporal coherence |
| **Scheduler** | beta57 | beta57 | Calibrated MiniMax curve |
| **Denoise** | 1.00 | *0.18 – 0.20** | **Do not exceed 0.25** to avoid face morphing |
| **Frame Rate** | 24 fps | 24 fps | Standard cinema frame rate |
| **Sharpening** | None | Radius: 1, Sigma: 0.8, Strength: 0.35 | Sub-second filmic enhancement |
---
## 💡 5. Best Practices & Pro Tips
* **Preserving Facial Likeness:** Keep Pass 2 denoise strictly between 0.16 and 0.20. A denoise of 0.50 will rewrite facial features and destroy identity.
* **Audio Synchronization:** The Audio VAE synthesizes a synchronized soundscape directly from the multimodal prompt. Use headphones to preview ambient audio and spatial sound effects.
* **Prompting Secret:** Let H3 Promptor do the heavy lifting. Avoid pasting giant walls of raw tags; instead, describe natural camera movement (*"slow tracking shot"*, "medium close-up") and physical dynamics (*"hair blowing gently in the wind"*).
---
## 📥 6. Installation & Files
1. Place the model checkpoint in:
ComfyUI/models/diffusion_models/H3_ErosLite.safetensors
2. Place the CLIP model in:
ComfyUI/models/clip/qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors
3. Place the VAEs in:
ComfyUI/models/vae/minimax_h3_video_vae_fp16.safetensors
ComfyUI/models/vae/minimax_h3_audio_vae_fp32.safetensors
4. Drag and drop the workflow file H3_ErosLite_2Pass_Clean_Workflow.json into ComfyUI.