Z-Image + LongCat Video Long-Form Generation Workflow: Complete Guide

Keywords: z-image longcat video generation, z-image long-form video, z-image comfyui video workflow, z-image longcat integration

In 2026, AI video generation is breaking past the "5-second clip" barrier. LongCat Video by Meituan—a 13.6B parameter open-source video generation foundation model—is redefining long-form video with its unified text-to-video, image-to-video, and video continuation capabilities. This guide covers the complete workflow of integrating Z-Image with LongCat Video in ComfyUI, from image generation to ultra-long video creation.

LongCat Video's Core Innovations

LongCat Video is the first truly unified long-form video generation model. Unlike earlier systems requiring separate models for different tasks, LongCat handles everything—text-to-video (T2V), image-to-video (I2V), and video continuation—with a single architecture.

Key Technical Breakthroughs

• Unified Video Generation Architecture: One model natively supports all three video generation tasks
• Video Continuation Pretraining: The key innovation enabling long, stable video generation—the model learns to understand "how to continue a video" from the start
• Coarse-to-Fine Generation: Gradually refines frames across time and space for efficient inference
• Block Sparse Attention: Major performance boost for high-resolution videos
• Multi-Reward RLHF (GRPO): Multiple reward signals optimize coherence, realism, and alignment

Why Z-Image + LongCat Video?

Combining Z-Image's image generation strength with LongCat's video capability creates a 1+1 > 2 effect:

Capability	Z-Image	LongCat Video	Combined
Image Generation	✅ High quality	❌ Not supported	Z-Image generates quality keyframes first
Text-to-Video	❌ Not supported	✅ Native	Z-Image description → LongCat generates video
Image-to-Video	❌ Not supported	✅ Native	Z-Image image → LongCat animates it
Video Continuation	❌ Not supported	✅ Native	Extend video infinitely
LoRA Style Control	✅	✅ (distill LoRA)	Unified visual style from image to video

Environment Setup

1. ComfyUI Foundation

Ensure ComfyUI is installed and running Z-Image models properly.

# Update ComfyUI and all custom nodes
# ComfyUI Manager → Update All

2. Install WanVideo Wrapper

LongCat Video is built on the Wan 2.2 architecture and requires Kijai's WanVideoWrapper custom node:

cd ComfyUI/custom_nodes/
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper.git

3. Download LongCat Models

Download from Kijai's HuggingFace repository:

Model	VRAM Required	File
FP8 Quantized	12-16 GB	LongCat_TI2V_comfy_fp8_e4m3fn_scaled_KJ.safetensors
BF16 Full Precision	24 GB+	LongCat_TI2V_comfy_bf16.safetensors

Place in ComfyUI/models/diffusion_models/.

4. Download LoRA Models

LoRA	Purpose	File
Refinement LoRA	Quality video in just 12 steps	LongCat_refinement_lora_rank128_bf16.safetensors
Distill LoRA	Speed up generation	LongCat_distill_lora_alpha64_bf16.safetensors

Place in ComfyUI/models/loras/.

5. Prepare Base Models

LongCat is based on Wan 2.2 TI2V, requiring these base models:

• wan_5b_TI2V_model → diffusion_models/
• wan2.1vae → vae/
• umt5-xxl → text_encoders/

Complete Z-Image → LongCat Video Workflow

Step 1: Generate Keyframes with Z-Image

Workflow node chain:
1. Z-Image Checkpoint Loader → Load Z-Image model
2. CLIP Text Encode → Enter detailed prompt
3. Z-Image KSampler → Generate high-quality image
4. Save Image → Save as keyframe

Prompt tip: Generated keyframes should include clear scene descriptions and compositional elements so LongCat understands the animation intent.

Step 2: Load LongCat Workflow

Drag ComfyUI/custom_nodes/ComfyUI-WanVideoWrapper/example_workflows/LongCat_TI2V_example_01.json into ComfyUI.

Key node groups:

1. Load Image — Input Z-Image keyframe
2. WanVideo Lora Select — Load refinement and/or distill LoRA
3. WanVideo VAE Loader — Load wan2.1vae
4. Wan Video Model Loader — Load LongCat model (select attention mode: flash att / sdpa / sage_atten)
5. Prompt — Enter detailed positive and negative prompts

Step 3: Key Parameters

Parameter	Recommended Value	Notes
CFG	1	LongCat recommends CFG=1
Shift	12	Schedule shift
Scheduler	LongCat distill Euler	Use distill for speed
Steps	12 (with refinement LoRA)	Refinement LoRA needs only 12 steps
Frames	81 (default)	Approximately 3-5 seconds per segment

Step 4: Extend to Long Video

The example workflow contains 5 identical groups. To extend video length:

1. Go to the last group
2. Detach the ImageBatchExtendWithOverlap and GetImageSizeAndCount nodes
3. Copy any group and paste after the last group
4. Add a new Reroute node
5. Connect: Previous group's ImageBatchExtendWithOverlap output → Reroute → New group's ImageBatchExtendWithOverlap input
6. Reattach: New last group's ImageBatchExtendWithOverlap → GetImageSizeAndCount

Repeat for arbitrary video lengths. Note: longer videos increase inference time.

Low VRAM Optimization

For 12-16 GB VRAM:

• Use FP8 quantized model
• Enable WanVideoBlockSwap node (block-swapping technique for lower VRAM)
• Reduce output resolution (720p instead of 1080p)
• Reduce frames per segment (41 instead of 81)

Practical Scenarios

Scenario 1: Product Demo Video

Z-Image generates high-quality product shot → LongCat creates 360° rotation animation
→ Video continuation adds multi-angle shots → Compose full product demo

Scenario 2: Dynamic Storyboard

Z-Image generates storyboard keyframes → LongCat converts each to short clip
→ Video continuation links all segments → Complete storyboard animation

Scenario 3: Social Media Shorts

Z-Image generates brand visuals → LongCat creates dynamic background animations
→ Overlay titles and transitions → Publish to social platforms

Troubleshooting

Unstable Quality or Flickering

• Reduce CFG to 1.0
• Ensure refinement LoRA is loaded
• Increase steps (16-20)
• Verify base Wan 2.2 models are intact

Color Drift in Long Videos

LongCat's pretraining significantly reduces color drift. If issues persist:

• Increase overlap frames in ImageBatchExtendWithOverlap
• Keep prompts consistent across continuation segments

Insufficient VRAM

• Switch to FP8 model
• Enable WanVideoBlockSwap
• Reduce resolution
• Decrease frames per segment

Summary

The Z-Image + LongCat Video combination opens up unprecedented possibilities for AI video creation:

• 🎬 From static images to dynamic video: Z-Image's high-quality outputs become perfect LongCat inputs
• 📏 Ultra-long video generation: Break through the 5-second limit with minute-long coherent videos
• 🎨 Style consistency: LoRA ensures visual style flows seamlessly from image to video
• 🚀 Low barrier to entry: FP8 quantization + block swap makes it run on 12GB VRAM

Whether you're a content creator, e-commerce operator, or brand designer, this Z-Image → LongCat Video workflow will dramatically boost your AI video creation efficiency.