ComfyUI provisioning manifest: Qwen-Image-Edit-2511 (FP8, 24GB card)

qwen-image-edit-2511-comfyui.yaml

# ComfyUI provisioner config: Qwen-Image-Edit-2511 (FP8, single 24GB card)
#
# Schema for the /opt/provisioner package shipped in vastai/comfy:*-auto images.
# Note this is NOT the same schema as the manifest used by the older
# /opt/instance-tools/bin/provisioner: models take repo/file/dest, not a url.
#
# Passed via COMFYUI_CONFIG_B64 so there is no gist CDN cache to fight.
version: 1

models:
  # VAE (250 MB)
  - source: huggingface
    repo: Comfy-Org/Qwen-Image_ComfyUI
    file: split_files/vae/qwen_image_vae.safetensors
    dest: vae/qwen_image_vae.safetensors

  # Text encoder: Qwen2.5-VL-7B FP8 scaled (9.38 GB). A VLM, it sees the image.
  - source: huggingface
    repo: Comfy-Org/Qwen-Image_ComfyUI
    file: split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors
    dest: text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors

  # Lightning distill LoRAs: 20-30 steps down to 8 or 4 (0.85 GB each)
  - source: huggingface
    repo: lightx2v/Qwen-Image-Edit-2511-Lightning
    file: Qwen-Image-Edit-2511-Lightning-8steps-V1.0-bf16.safetensors
    dest: loras/Qwen-Image-Edit-2511-Lightning-8steps-V1.0-bf16.safetensors

  - source: huggingface
    repo: lightx2v/Qwen-Image-Edit-2511-Lightning
    file: Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors
    dest: loras/Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors

  # Diffusion model: 20B MMDiT, FP8 e4m3fn scaled (20.49 GB).
  # Scaled, not plain e4m3fn: same size, better quality.
  - source: huggingface
    repo: lightx2v/Qwen-Image-Edit-2511-Lightning
    file: qwen_image_edit_2511_fp8_e4m3fn_scaled.safetensors
    dest: diffusion_models/qwen_image_edit_2511_fp8_e4m3fn_scaled.safetensors

qwen-image-edit-2511.yaml

# ComfyUI provisioning: Qwen-Image-Edit-2511
#
# Consumed by the vast.ai provisioner via the PROVISIONING_MANIFEST env var.
# Every phase is content-hash checked and downloads skip existing files, so a
# re-run or a retry after an interrupted pull costs nothing.
#
# Qwen image-edit releases are date-stamped, not semver: Edit -> 2509 -> 2511.
# There is no "3.0". 2511 (Nov 2025) is the current release.
#
# Models are the FP8 set (31.8 GB), sized for a single 24 GB card. The full
# BF16 model is 57.7 GB and needs 2x48 GB, so it is deliberately not used here.
#
# VRAM note: DiT (20.5 GB) + text encoder (9.4 GB) exceeds 24 GB. ComfyUI runs
# the encoder, evicts it, then loads the DiT. That is expected. Do not try to
# keep both resident.
#
# Custom nodes are deliberately omitted. Anonymous git clones from datacenter
# IPs hit GitHub rate limits ("could not read Username for github.com"), and
# the provisioner runs git_repos BEFORE downloads, so a failing clone phase
# blocks the model pull entirely. None are needed for Qwen-Image-Edit. Install
# extras from the ComfyUI-Manager UI after boot instead.
#
# Requires ~120 GB disk: 15 GB base image, 31.8 GB models, ~21 GB transient
# download space, rest for output.

version: 1

settings:
  venv: "/venv/main"
  concurrency:
    # Downloads land in a tmpdir then move. Same filesystem, so the move is a
    # rename, but N parallel pulls hold N files at once. 2 keeps the transient
    # peak near 21 GB instead of 31 GB while staying reasonably fast.
    hf_downloads: 2

downloads:
  # --- VAE (250 MB) ---
  - url: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/vae/qwen_image_vae.safetensors
    dest: "${WORKSPACE:-/workspace}/ComfyUI/models/vae/qwen_image_vae.safetensors"

  # --- Text encoder: Qwen2.5-VL-7B, FP8 scaled (9.38 GB).
  #     This is a VLM, it actually sees the input image. ---
  - url: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors
    dest: "${WORKSPACE:-/workspace}/ComfyUI/models/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors"

  # --- Lightning distill LoRAs: cut 20-30 steps to 4 or 8.
  #     Use while iterating, drop for finals (0.85 GB each). ---
  - url: https://huggingface.co/lightx2v/Qwen-Image-Edit-2511-Lightning/resolve/main/Qwen-Image-Edit-2511-Lightning-8steps-V1.0-bf16.safetensors
    dest: "${WORKSPACE:-/workspace}/ComfyUI/models/loras/Qwen-Image-Edit-2511-Lightning-8steps-V1.0-bf16.safetensors"

  - url: https://huggingface.co/lightx2v/Qwen-Image-Edit-2511-Lightning/resolve/main/Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors
    dest: "${WORKSPACE:-/workspace}/ComfyUI/models/loras/Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors"

  # --- Diffusion model: 20B MMDiT, FP8 e4m3fn scaled (20.49 GB).
  #     Scaled, not plain e4m3fn: same size, better quality. ---
  - url: https://huggingface.co/lightx2v/Qwen-Image-Edit-2511-Lightning/resolve/main/qwen_image_edit_2511_fp8_e4m3fn_scaled.safetensors
    dest: "${WORKSPACE:-/workspace}/ComfyUI/models/diffusion_models/qwen_image_edit_2511_fp8_e4m3fn_scaled.safetensors"