ARSADR/colab-model-weights

Colab Model Weights & Workflows Vault

Centralized, high-throughput model weight repository and workflow library curated for Google Colab GPU runtime environments (ComfyUI, vLLM, llama-server).

Optimized for rapid zero-egress downloads, native FP8 / INT8-convrot / NVFP4 architectures (NVIDIA Blackwell sm_120 and Ada sm_89), and automated pipeline staging.

---

📦 Directory Structure & Inventory

.
├── bbox/                                       # Detection & crop models
│   └── face_yolov8m.pt                         (49.6 MB)  YOLOv8 face detector for crop/alignment
├── checkpoints/                                # Integrated checkpoints
│   ├── Qwen-Rapid-AIO-NSFW-v23.safetensors     (26.5 GB)  Qwen-Image-2.1 All-In-One pipeline
│   └── v1-5-pruned-emaonly.safetensors         (3.97 GB)  SD 1.5 base checkpoint (mask/control)
├── diffusion_models/                           # Primary DiT & diffusion backbones
│   ├── fasth3_omniref_8step_pruned_int8_convrot.safetensors            (20.7 GB)  FastH3 OmniRef 8-step DiT (INT8-convrot)
│   ├── Minimax-h3_Singularity_ref2va_v2_pruned_int8.safetensors        (19.5 GB)  Singularity Ref2VA v2 DiT (INT8)
│   ├── Minimax-h3_Singularity_x_10Eros-Max_PRISM_int8.safetensors      (18.9 GB)  Singularity x 10Eros-Max PRISM DiT (INT8)
│   ├── minimax_h3_fused_refdelta_r1024_turbo8_mystic07_int8_convrot.safetensors (19.5 GB) Fused RefDelta Turbo8 Mystic (INT8-convrot)
│   ├── Qwen-Image-2.1-Turbo-Uncensored-Q8_0.gguf                      (8.11 GB)  Qwen-Image-2.1 Turbo Uncensored (GGUF Q8_0)
│   └── qwen-image-2.1-turbo-Q8_0.gguf                                  (7.12 GB)  Qwen-Image-2.1 Turbo DiT (GGUF Q8_0)
├── latent_upscale_models/                      # 3D spatial/temporal upscalers
│   └── minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors          (0.64 GB)  MiniMax H3 3D-Conv Latent Upscaler (BF16)
├── llm/                                        # Large Language Model runtimes & projectors
│   ├── qwen3.8-27b/                            # Qwen 3.8 27B Vision-LLM suite
│   │   ├── Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf     (16.7 GB)  27B multimodal backbone (Q4_K_P)
│   │   ├── Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-FastMTP-32K.gguf (0.84 GB) Speculative decoding draft / MTP adapter
│   │   └── mmproj-Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-BF16.gguf (0.87 GB) Vision projector (BF16)
│   └── qwythos-9b/                             # Qwythos 9B 1M-context suite
│       ├── Qwythos-9B-Claude-Mythos-5-1M-MTP-Q8_0.gguf                (9.11 GB)  9B 1M-context MTP model (Q8_0)
│       └── mmproj-Qwythos-9B-Claude-Mythos-5-1M-F16.gguf              (0.86 GB)  Vision projector (F16)
├── loras/                                      # Specialized adaptation weights
│   └── minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors      (1.82 GB)  4-step Turbo acceleration LoRA (BF16)
├── refs/                                       # Multi-angle reference & condition assets
│   ├── ref_charsheet_cinema.png                (0.79 MB)  Cinematic wide character sheet
│   ├── ref_charsheet_dessert.png               (1.72 MB)  Pastry/dessert multi-angle reference
│   ├── ref_charsheet_rooftop.png               (1.55 MB)  Rooftop setting character sheet
│   ├── ref_portrait_letter.png                 (1.07 MB)  Close-up portrait reference
│   ├── ref_portrait_ponytail.png               (1.43 MB)  Hairstyle / identity reference
│   ├── ref_storyboard_fight.png                (0.53 MB)  Dynamic action storyboard reference
│   └── ref_images_minimax_h3.tar.gz            (7.09 MB)  Consolidated reference bundle
├── text_encoders/                              # Text & multimodal conditioners
│   ├── qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors                (14.6 GB)  Qwen3-VL 32B Heretic (NVFP4, sm_120 native)
│   ├── qwen3.5_9b_qwen_image_2.1_pe_i2i.int8_convrot.safetensors       (8.82 GB)  Qwen 3.5 9B Prompt Enhancer (INT8-convrot)
│   └── qwen_image_2.1_turbo_qwen3vl_8b_official.safetensors            (16.3 GB)  Official Qwen3-VL 8B Conditioner (BF16)
├── vae/                                        # Autoencoders
│   ├── minimax_h3_video_vae_int8_convrot.safetensors                   (2.95 GB)  MiniMax H3 Video VAE Decoder/Encoder (INT8-convrot)
│   ├── minimax_h3_audio_vae_fp32.safetensors                           (0.56 GB)  MiniMax H3 Audio VAE (FP32)
│   ├── qwen_image_2.1_turbo_vae_official.safetensors                   (0.63 GB)  Official Qwen-Image 2.1 VAE
│   └── qwen_image_2.1_vae_bf16.safetensors                             (0.63 GB)  Qwen-Image 2.1 VAE (BF16)
└── workflows/                                  # Declarative ComfyUI graph pipelines
    ├── character-swap/                         # Face relay & character swap pipelines
    │   ├── H3-Character-Swap-Relay-V2.api.json
    │   ├── H3-Character-Swap-Relay-V2.json
    │   ├── H3_Character_Swap_v1_Ref2VA_akatz.json
    │   ├── amao2001_H3_Character_Swap_v1_Ref2VA.json
    │   ├── amao2001_swap_mask.json
    │   ├── examples_README.md
    │   └── turbo8_adaptive_window.json
    ├── minimax-h3/                             # MiniMax H3 production video workflows
    │   ├── 05_ref2va_4step_sla.json
    │   ├── 06_ref2va_4step_sla_qwen38.json
    │   ├── 08_ref2va_native_nvfp4_textgen.json
    │   ├── 09_minimax_h3_all_in_one_director_g4.json
    │   ├── fastH3_ref2va.json
    │   ├── minimax_h3_i2v_fastH3_face_detail.json
    │   ├── minimax_h3_ref2va_fastH3_trim_latent_upscale.json
    │   └── variants/                           # FastH3 and Singularity specialized variants
    │       ├── 10eros_singularity_base1.json
    │       ├── 10eros_singularity_latent_upscale8.json
    │       ├── fast_h3_omni_base.json
    │       ├── fast_h3_omni_latent_upscale.json
    │       ├── singularity_v2_face_refine.json
    │       ├── singularity_v2_r2va.json
    │       └── README.md
    ├── qwen-image/                             # Qwen-Image 2.1 generation & editing
    │   ├── Qwen-Image-2.1-viggle-turbo-edit.json
    │   ├── Qwen-Rapid-AIO.json
    │   ├── qwen218StepsWithPrompt_v14.json
    │   ├── qwen21_image_edit_heretic_gguf.json
    │   ├── qwen21_image_edit_official.json
    │   ├── qwen21_image_edit_pottokao_2stage_standard_ui.json
    │   ├── qwen21_image_edit_pottokao_full_ui.json
    │   ├── qwen21_native_clean.json
    │   ├── qwen21_pottokao_authentic_recipe.json
    │   ├── qwen21_pottokao_exact_all_gguf.json
    │   ├── qwen21_pottokao_image_edit_all_gguf.json
    │   ├── qwen21_workflow_no_translate.json
    │   ├── qwen21turbo_uncensored.json
    │   └── viggle_v021_image_edit_ui.json
    ├── image_qwen_image_2_1_turbo_image_edit.json
    └── image_qwen_image_2_1_turbo_t2i.json

---

⚡ Quantization & Hardware Compatibility

Model Component	Format / Precision	Target Hardware	Notes
qwen3vl_32b_heretic_minimax_h3_nvfp4	NVFP4	NVIDIA Blackwell (sm_120 / G4)	Extreme throughput & low VRAM footprint on native Blackwell Tensor Cores.
fasth3_omniref_8step_pruned_int8_convrot	INT8-convrot	NVIDIA Ada (sm_89), Blackwell (sm_120), Hopper	8-step high-efficiency video DiT with rotational invariant convolution kernels.
minimax_h3_video_vae_int8_convrot	INT8-convrot	Universal CUDA	Low-overhead temporal video decoding.
Qwen-Image-2.1-Turbo-*.gguf	GGUF Q8_0	Universal CUDA / llama.cpp / ComfyUI-GGUF	High visual fidelity image generation with quantized weights.
Qwen3.8-27B-Uncensored-*	GGUF Q4_K_P + BF16 mmproj	NVIDIA L4 / A100 / Blackwell	Full multimodal context understanding with speculative decoding draft support.
minimax_h3_latent_upscaler_3d_conv_v1	BF16	NVIDIA Ada / Blackwell	3D spatio-temporal upscaling without tiling boundary artifacts.

---

🚀 Staging in Google Colab

Download directly into standard ComfyUI directories using the high-speed Hugging Face Hub client:

from huggingface_hub import hf_hub_download
import os

repo_id = "ARSADR/colab-model-weights"
comfy_root = "/content/ComfyUI"

# Example: Staging FastH3 Omni Base weights
downloads = [
    ("diffusion_models/fasth3_omniref_8step_pruned_int8_convrot.safetensors", "models/diffusion_models"),
    ("text_encoders/qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors", "models/text_encoders"),
    ("vae/minimax_h3_video_vae_int8_convrot.safetensors", "models/vae"),
    ("vae/minimax_h3_audio_vae_fp32.safetensors", "models/vae"),
    ("latent_upscale_models/minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors", "models/latent_upscale_models"),
]

for remote_path, local_subfolder in downloads:
    dest_dir = os.path.join(comfy_root, local_subfolder)
    os.makedirs(dest_dir, exist_ok=True)
    filename = os.path.basename(remote_path)
    target = os.path.join(dest_dir, filename)
    if not os.path.exists(target):
        print(f"Downloading {filename}...")
        hf_hub_download(
            repo_id=repo_id,
            filename=remote_path,
            local_dir=dest_dir,
            local_dir_use_symlinks=False
        )

For complete automated serving runbooks, consult the Colab CLI Playground runtime guides.