digitalsamba__claude-code-video-toolkit/docker/runpod-qwen-edit

RunPod Qwen-Edit Docker Image

Serverless GPU handler for AI-powered image editing using Qwen-Image-Edit-2511 with LightX2V acceleration.

This is Worker 1 in the video generation pipeline:

Reference Image -> [Qwen-Edit] -> Edited Frame -> [Wan I2V] -> Video Clip

Quick Start

Option A: Use Pre-built Public Image (Recommended)

ghcr.io/conalmullan/video-toolkit-qwen-edit:latest

Skip to Step 2: Deploy on RunPod below.

Option B: Build Your Own Image

cd docker/runpod-qwen-edit

# Build for linux/amd64 (required for RunPod)
docker buildx build --platform linux/amd64 -t ghcr.io/yourusername/video-toolkit-qwen-edit:latest --push .

Build takes ~45-60 minutes (downloads ~30GB of model weights).

Important: Make the image public

GHCR images are private by default. RunPod cannot pull private images.

1. Go to https://github.com/users/yourusername/packages/container/video-toolkit-qwen-edit/settings
2. Scroll to "Danger Zone"
3. Click "Change visibility" → Select "Public" → Confirm

Deploy on RunPod

1. Go to RunPod Serverless
2. Click New Endpoint
3. Configure:
   • Docker Image: ghcr.io/conalmullan/video-toolkit-qwen-edit:latest
   • GPU: L4 24GB or RTX 4090 (24GB+ VRAM required)
   • Max Workers: 1 (scale up as needed)
   • Idle Timeout: 5 seconds (fast scale-down)
   • Execution Timeout: 120 seconds (2 min max per job)
4. Copy the Endpoint ID for your .env file

Configure Local Tool

Add to your .env:

RUNPOD_API_KEY=your_api_key_here
RUNPOD_QWEN_EDIT_ENDPOINT_ID=your_endpoint_id_here

Use It

# Generate edited frame (via orchestration tool - coming soon)
python tools/videogen.py edit \
    --image reference.png \
    --prompt "Change background to modern office" \
    --output edited.png

Image Details

Property	Value
Base	nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04
Size	~25GB
Cold Start	~60 seconds
Python	3.11
PyTorch	2.5.1 + CUDA 12.4

Pre-baked Components

• Qwen-Image-Edit-2511 base model (~20GB)
• LightX2V Lightning FP8 weights (~10GB)
  • 4-step inference (vs 40 native)
  • FP8 quantization (50% VRAM reduction)
• Flash Attention 3 for optimized inference
• FFmpeg for image processing
• RunPod SDK

API Reference

Edit Operation

Edit an image based on text prompt while preserving subject identity.

Input:

{
    "input": {
        "image_base64": "<base64 encoded image>",
        "prompt": "Change the background to a modern office",
        "negative_prompt": "blurry, distorted",
        "num_inference_steps": 8,
        "guidance_scale": 1.0,
        "seed": 42,
        "use_fp8": true,
        "r2": {
            "endpoint_url": "https://...",
            "access_key_id": "...",
            "secret_access_key": "...",
            "bucket_name": "..."
        }
    }
}

Parameter	Required	Default	Description
image_base64	Yes	-	Base64 encoded input image
prompt	Yes	-	Edit instruction
negative_prompt	No	""	Things to avoid
num_inference_steps	No	8 (FP8), 4 (LoRA)	Diffusion steps
guidance_scale	No	1.0	CFG scale
seed	No	random	Random seed for reproducibility
use_fp8	No	true	Use FP8 quantization (lower VRAM)
auto_resize	No	true	Auto-resize for optimal processing
r2	No	-	R2 config for result upload

Output:

{
    "success": true,
    "edited_image_base64": "<base64 encoded result>",
    "seed": 42,
    "inference_time_ms": 1800,
    "image_size": [1024, 1024],
    "num_inference_steps": 8,
    "use_fp8": true
}

If R2 config provided:

{
    "success": true,
    "edited_image_base64": "<base64>",
    "output_url": "https://r2.example.com/...",
    "r2_key": "qwen-edit/results/abc123.png",
    ...
}

Example Prompts

Background changes:

• "Change the background to a modern office"
• "Replace background with outdoor park scene"
• "Add a professional studio backdrop"

Pose/expression adjustments:

• "Change pose to speaking gesture"
• "Add subtle smile"
• "Subject looking slightly left"

Lighting/style:

• "Add professional lighting"
• "Apply warm color grading"
• "Make lighting more dramatic"

GPU Memory Requirements

GPU VRAM	FP8 Mode	BF16 Mode	Notes
24GB	Yes	Yes*	*May be tight
48GB	Yes	Yes	Recommended
80GB	Yes	Yes	Fastest

Recommendation: Use L4, RTX 4090, or A6000 (24GB+) with FP8 mode.

Performance

Using L4 24GB (~$0.34/hr):

Mode	Steps	Time	Cost/Edit
FP8 Lightning	8	~2s	~$0.0002
LoRA Lightning	4	~1.5s	~$0.00014
Native (no accel)	40	~15s	~$0.0014

LightX2V acceleration provides ~10x speedup and ~50% VRAM reduction.

Local Testing

Test the image locally with NVIDIA GPU:

# Build
docker build -t qwen-edit-test .

# Run interactive shell
docker run --gpus all -it qwen-edit-test /bin/bash

# Inside container, test GPU
python3 -c "import torch; print(f'CUDA: {torch.cuda.is_available()}')"

# Test LightX2V
python3 -c "from lightx2v import LightX2VPipeline; print('LightX2V OK')"

# Test model paths
python3 -c "from pathlib import Path; print(f'Model: {Path(\"/models/qwen-edit\").exists()}')"

Troubleshooting

"CUDA out of memory"

1. Ensure use_fp8: true is set (default)
2. Use a GPU with 24GB+ VRAM
3. Enable auto_resize to scale down large images

"Generation failed - no output file"

Check the prompt syntax. Qwen-Image-Edit works best with:

• Clear, specific instructions
• Focus on ONE change at a time
• Avoid contradictory instructions

Cold start is slow (~60s)

First request after idle loads the full model. Options:

• Set longer idle timeout
• Use "always on" worker for frequent usage
• Pre-warm with a test request

Identity drift in chained edits

This is a known challenge. Mitigations:

• Use consistent seed across edits
• Keep prompts focused on single changes
• Consider IP-Adapter for explicit identity conditioning (future)

Pipeline Integration

This worker is designed for the video generation pipeline:

# Pseudo-code for pipeline orchestration
def generate_video_segment(reference_image, scenes):
    current_frame = reference_image

    for scene in scenes:
        # Worker 1: Edit frame
        edited = call_qwen_edit(current_frame, scene["edit_prompt"])

        # Worker 2: Animate (separate endpoint)
        video = call_wan_i2v(edited, scene["motion_prompt"])

        # Chain: last frame becomes next input
        current_frame = video["last_frame"]

    return concat_videos(videos)

See .ai_dev/video-generation-pipeline.md for full pipeline documentation.

Related Workers

Worker	Purpose	Status
runpod-qwen-edit	Frame editing (this)	Phase 1
runpod-wan-i2v	Image-to-video	Phase 2 (planned)
runpod-propainter	Watermark removal	Deployed
runpod-animate	SVD animation	Deployed