Extra-long ComfyUI runs: VRAM and overlap

Browser fallback: if extra-long concatenation runs out of VRAM, or the joins drift after a few segments, skip ComfyUI and generate in the browser with the hosted image-to-video workflow — no local GPU.

Extra-long ComfyUI runs: VRAM, overlap drift, then the browser

Prefer not to install anything? The generator below is the same hosted LongCat generator as on the home page (image-to-video and text-to-video) — you can run it right here:

Introduction: Breaking the Duration Barrier

LongCat Video introduces a revolutionary approach to AI video generation - the ability to create extra long videos that extend far beyond the traditional 5-10 second limitations. Through ComfyUI's powerful workflow system and LongCat's unified architecture, you can now generate minutes-long coherent videos using a sophisticated concatenation technique.

If you only need to know how long a single online generation can be — frame limits, fps and credit cost — see How Long Can LongCat Video Generate?.

This comprehensive guide will walk you through the complete workflow for generating extra long videos, from basic setup to advanced prompt engineering techniques. Whether you're creating YouTube content, product demonstrations, or cinematic sequences, you'll learn how to maximize LongCat Video's potential for extended video generation.

TL;DR

LongCat Video's extra long generation workflow in ComfyUI enables unlimited video duration through intelligent concatenation. This tutorial covers:

• ✅ Unified Architecture: Single model handles text-to-video, image-to-video, and video continuation
• ✅ Concatenation Workflow: Chain multiple segments with overlap for seamless long videos
• ✅ VRAM Optimization: Block swap settings for 12GB-24GB GPUs
• ✅ Modular Design: Easy-to-follow node groups for complex workflows
• ✅ Practical Examples: Real-world tests with different settings and resolutions
• ✅ Multi-Prompt System: Create narrative sequences with 4+ connected prompts

Who This Guide Is For:

• 🎥 ComfyUI users wanting to extend video duration
• 📹 Content creators needing long-form AI video
• 🎬 Filmmakers exploring AI cinematography
• 💻 Developers building custom video workflows

Expected Outcome: After reading this guide, you'll be able to generate 30+ second coherent videos by chaining multiple segments, with the knowledge to extend even further.

Reading Time: 12 minutes | Difficulty: Intermediate

💡 Key Takeaways

• 🚀 LongCat Video uses a unified architecture supporting text-to-video, image-to-video, and video continuation in one model
• 🎯 Extra long generation works through concatenation with configurable overlap (13 frames default)
• 💰 FP8 model version reduces VRAM requirements while maintaining quality
• ⚡ Block swap settings let you run on 12GB VRAM (40 blocks) or optimize for 24GB (11 blocks)
• 🎨 Modular workflow design makes complex multi-segment generation easy to manage
• 📊 Settings like CFG, shift, and steps dramatically affect output quality and style
• 🔊 Each segment can have unique prompts, enabling narrative progression
• 🎬 Resolution up to 720p with 93 frames per segment recommended
• 💾 Shift values of 12-24 and steps of 12-16 provide best balance of quality and speed
• 🤝 Seamless transitions achieved through overlap and shared latent space

1. Understanding LongCat Video's Unified Architecture

1.1 What Makes LongCat Unique

LongCat Video stands out from other AI video generators through its unified architecture - a single model that natively supports three core generation modes without requiring separate systems:

Text-to-Video: Generate videos directly from text descriptions with sophisticated scene understanding and temporal coherence.

Image-to-Video: Animate static images with natural motion while preserving the original composition and character identity.

Video Continuation: Seamlessly extend existing videos by understanding motion patterns and maintaining visual consistency.

Traditional video generation models require different architectures or fine-tuned versions for each task. LongCat's unified approach means:

• Consistent visual quality across all modes
• Simplified workflow management
• Efficient VRAM usage
• Faster iteration and experimentation

1.2 The Concatenation Principle

The key to generating extra long videos lies in LongCat's concatenation workflow. Here's how it works:

1. Initial Generation: Create the first segment (5-8 seconds) from text or image
2. Overlap Extraction: Take the last 13 frames of the previous segment
3. Continuation: Generate next segment using overlap as reference
4. Seamless Merge: Combine segments with shared frames for smooth transitions
5. Repeat: Chain as many segments as needed for desired duration

Why Overlap Matters: The 13-frame overlap ensures visual and motion continuity between segments. Without overlap, you'd see jarring cuts and inconsistent character appearances. With proper overlap, the model maintains color consistency, character identity, and motion direction across the entire sequence.

1.3 Key Features for Extra Long Generation

Frame-Based Control: Each segment generates 93 frames by default at 30fps, yielding approximately 3 seconds of video. Multiply segments for longer duration.

Consistent Color and Identity: LongCat's 3D attention mechanism and Rope positional encoding prevent the "drift" problem that plagues extended generation in other models.

Flexible Resolution: Support for resolutions from 480p up to 720p, with higher resolutions producing noticeably better results in building details, character features, and environmental coherence.

2. Setting Up the ComfyUI Workflow

2.1 Model Loader Configuration

The foundation of your workflow starts with proper model loading:

Model Selection: Use the FP8 version of LongCat Video for optimal VRAM efficiency. The FP8 quantization reduces memory requirements by approximately 40% compared to the full precision model.

Critical Setting - Base Precision: Always set base precision to BF16 (Brain Float 16). Other options like FP16 will produce degraded results with color artifacts and inconsistent generation.

Attention Mode: Best results come from Sage Attention if available. If you're new to ComfyUI or don't have advanced attention mechanisms installed, use SDPA (Scaled Dot-Product Attention) as a reliable fallback.

2.2 LoRA Integration

LongCat Video includes a distillation LoRA for improved efficiency:

LoRA Strength: Default strength of 1.0 works well, but experimentation shows good results between 0.7-1.0. Lower strengths can sometimes produce more creative variations.

Multi-Loader Setup: Even though there's currently one LoRA, using a multi-loader node prepares your workflow for future LoRAs without restructuring.

2.3 Block Swap and VRAM Management

The block swap feature is crucial for managing VRAM consumption:

VRAM Tiers:

• 11 Blocks to Swap: Requires ~24GB VRAM (RTX 3090/4090/5090)
• 20-30 Blocks: Requires ~16GB VRAM (RTX 4080)
• 40 Blocks: Runs on ~12GB VRAM (RTX 4070 Ti)

Performance Trade-off: More blocks swapped = lower VRAM usage but slower generation. Find the sweet spot for your hardware.

Experimental Options: Advanced users can experiment with cache attachment and other optimization flags, though default settings produce the most reliable results.

2.4 Torch Compile Option

The workflow includes optional Torch Compile acceleration:

• Enables PyTorch's JIT compilation for faster execution
• Can be disabled if not supported on your system
• Provides 10-15% speed improvement when working

3. Configuring Global Settings

3.1 Core Parameters

The workflow uses centralized settings that apply to all K-sampler nodes:

Resolution Settings:

• Width: 1024 or 1280 recommended
• Height: 512 or 640 recommended
• Aspect ratio: 2:1 works best for cinematic shots

Temporal Settings:

• Frames: 93 (default, ~3 seconds at 30fps)
• Overlap: 13 frames (ensures smooth transitions)

Sampling Settings:

• Steps: 12-16 (12 with LoRA, 50 without)
• CFG: 1.0 with LoRA, 4.0 without
• Shift: 12-24 (higher values = more stylized)

3.2 Understanding Shift Values

The shift parameter significantly affects visual style:

Shift = 12: Realistic, grounded appearance with natural motion Shift = 18: Slightly enhanced stylization, more defined shapes Shift = 24: Stronger artistic interpretation, more cinematic feel

Test Results: Side-by-side comparisons show shift=24 with 14 steps produces the most building-like buildings and coherent environmental details, while shift=12 offers faster generation with acceptable quality.

3.3 Resolution Impact

Higher resolutions dramatically improve output quality:

640x480: Fast generation, acceptable for testing 1024x512: Good balance of speed and quality 1280x640: Noticeably improved detail in buildings, faces, and environments (recommended)

Note: Higher resolutions require proportionally more VRAM. Adjust block swap accordingly.

4. Multi-Prompt System for Narrative Sequences

4.1 Prompt Structure

The modular workflow supports multiple sequential prompts for narrative progression:

Prompt 1: Establish the scene and main character Prompt 2: Introduce movement or environment change Prompt 3: Add action or interaction Prompt 4: Climax or conclusion

Each prompt generates a separate video segment that concatenates with the previous one.

4.2 Practical Example: The Rodent Rider

Here's a real workflow example from the video tutorial:

Prompt 1 (Scene Establishment):

A rodent riding a sleek, powerful red and gold tech hover bike
through an alien city center with tall, gothic, and metallic
buildings. Futuristic film style scene. The rodent wears a black
leather biker jacket with massive black boots and silver buckles.

Prompt 2 (Environment Transition):

A rodent riding a space bike through an alien forest from the
future. The trees are very tall and it feels like autumn.

Prompt 3 (Character Interaction):

A rodent riding a space bike through an alien forest from the
future. He smiles and waves quickly at the camera.

Prompt 4 (Dramatic Conclusion):

A leather jacket-wearing rodent dude rides a space bike which
suddenly powers up and achieves liftoff, flying up from the road
and into outer space. The darkness of the void-like cosmos around
him, stars twinkling as the cosmic rider starts on his true journey.

4.3 Prompt Engineering Best Practices

Continuity Elements: Reference consistent details across prompts (character clothing, bike appearance, weather conditions)

Progressive Description: Each prompt should naturally flow from the previous one while introducing new elements

Descriptive Specificity: Include visual details like colors, materials, lighting, and mood for better results

Action Clarity: Use clear verbs and motion descriptions (riding, waving, lifting off) for coherent animation

5. K-Sampler Workflow Design

5.1 Text-to-Video Sampler

The first sampler generates from text:

Components:

• WAN Video Empty Embeds: Creates base latent space
• LongCat Distill Scheduler: Specialized denoising schedule
• K-Sampler: Performs iterative denoising with prompt guidance

Settings Reference: Uses global width, height, frames, steps, CFG, and shift values

Output: Initial video segment serving as foundation for continuation

5.2 Image-to-Video Sampler

Alternative starting point using a reference image:

Additional Components:

• Image Batch Extend with Overlap: Prepares image for video encoding
• WAN Video Encode: Converts image to video latent space
• Extra Latent Input: Conditions generation on reference image

Use Case: When you have a specific starting frame or character design you want to animate

Workflow Toggle: Simple switch node lets you choose text-to-video or image-to-video without restructuring

5.3 Continuation Samplers

Subsequent samplers extend the video:

Video Input: Previous segment feeds into Image Batch Extend Overlap Processing: Extracts last 13 frames for continuity Encoding: Converts frames to latent space for conditioning Generation: Creates next segment with seamless transition Output: Extended video passed to next sampler or final output

Modular Replication: Copy the continuation group for as many segments as needed (prompt 2, 3, 4, 5...)

5.4 Critical Updates When Duplicating

When creating additional segments, remember to update:

✅ Prompt Number: Change to prompt_2, prompt_3, etc. ✅ Video In/Out: Update video_1 → video_2 → video_3 progression ✅ Seed Value: Increment or randomize for variation (e.g., 420 → 421 → 422)

Tip: Missing these updates will cause errors or identical generation instead of progression.

6. Real-World Testing Results

6.1 Baseline Test (12 Steps, Shift 12)

Settings:

• Steps: 12
• Shift: 12
• Resolution: 1024x512
• Total Duration: ~12 seconds (4 segments)

Results: Good quality generation with acceptable building structure, smooth forest transition, somewhat inconsistent hand/finger rendering during wave, successful takeoff sequence.

6.2 Increased Shift (12 Steps, Shift 18)

Change: Shift value increased to 18

Observed Differences:

• More defined building architecture
• Slightly different bike design variations
• Improved environmental coherence
• Wave gesture changed (different finger count/motion)

Conclusion: Higher shift values tend toward more stylized, architecturally consistent results.

6.3 Maximum Shift (12 Steps, Shift 24)

Settings: Shift increased to 24

Results:

• Most building-like buildings across all tests
• Trees appear more stable and realistic
• Double wave gesture (character waves twice)
• Different takeoff animation (middle of scene)

Trade-off: More stylization but potentially less literal prompt following.

6.4 Increased Steps (14-16 Steps, Various Shifts)

14 Steps + Shift 24:

• Further improved building quality
• Better wave gesture with near-correct finger count
• Smoother motion overall

16 Steps + Shift 24:

• Best overall quality in testing
• Buildings maintain consistency
• Takeoff sequence happens mid-scene (creative interpretation)

16 Steps + Shift 12:

• More grounded realism
• Excellent takeoff sequence
• Better prompt adherence than high shift values

6.5 High Resolution Test (1280x640)

Settings:

• Resolution: 1280x640
• Steps: 14
• Shift: 24

Results:

• Dramatically improved building detail and realism
• Character features more refined (hair, clothing detail visible)
• Correct finger count in wave gesture
• Best overall visual quality across all tests

Recommendation: Use higher resolutions when VRAM allows - the quality improvement is significant.

7. Final Video Assembly

7.1 Video Concatenation Node

The final workflow group assembles all segments:

Input: All generated video segments (video_1, video_2, video_3, video_4...) Process: Extend Video node chains them with overlap consideration Output: Single unified video file with seamless transitions

7.2 Quality Control

Check Points:

• Verify smooth transitions between segments
• Confirm character/environment consistency
• Review motion continuity
• Check color grading uniformity

Common Issues:

• Jarring cuts (increase overlap frames)
• Character appearance drift (refine prompts for consistency)
• Motion discontinuity (adjust shift/CFG values)

❓ Frequently Asked Questions

Q1: What's the maximum video length I can generate?

A: Theoretically unlimited! Each segment is approximately 3 seconds (93 frames at 30fps). You can chain as many segments as you want by duplicating the continuation sampler groups. The video tutorial demonstrates a 4-segment workflow creating ~12 seconds, but you can easily extend to 10+ segments for 30+ second videos.

Practical Limitation: Generation time increases linearly with segments. A 4-segment video might take 15-25 minutes on RTX 4090, so a 10-segment video would take 40-60 minutes.

Q2: Can I run this workflow on 8GB VRAM?

A: Not recommended for the full FP8 model. The minimum viable VRAM is approximately 12GB with block swap set to 40. For 8GB cards:

• Consider cloud generation alternatives
• Wait for further model optimizations
• Use lower resolutions (480p) if willing to compromise quality

Q3: Should I use the LoRA or not?

A: Yes, highly recommended. The distillation LoRA allows:

• Reduced steps (12-16 instead of 50)
• Lower CFG value (1.0 instead of 4.0)
• Faster generation with comparable quality
• More consistent results

Without LoRA, you'll need 50 steps and CFG of 4.0, significantly increasing generation time.

Q4: Why are my transitions not smooth between segments?

A: Common causes and solutions:

Issue: Overlap frames set too low Solution: Keep default 13 frames or increase to 17-21

Issue: Prompts too different between segments Solution: Maintain consistent descriptive elements (character appearance, environment, lighting)

Issue: Different shift/CFG values between samplers Solution: Use global settings node to ensure consistency

Issue: Seed values causing divergence Solution: Use incremental seeds (420, 421, 422) rather than random

Q5: What's the best resolution for quality vs speed?

A: Depends on your use case:

Recommendation: Test at 1024x512, finalize at 1280x640 for best balance.

Q6: How do shift values affect the output?

A: Shift controls stylization vs realism:

Shift 12: Realistic, natural motion, literal prompt interpretation Shift 18: Slightly stylized, enhanced shapes, good detail Shift 24: Artistic interpretation, best architectural coherence, creative variations

Best Practice: Try shift=12 for realistic content, shift=24 for cinematic/artistic content.

Q7: Can I switch between text-to-video and image-to-video mid-generation?

A: No, the starting mode (text or image) applies to the first segment only. All subsequent segments use video continuation regardless of initial mode. However, you could:

• Generate first segment with text-to-video
• Use that output as reference image
• Start a new workflow with image-to-video
• Continue from there

Q8: Why do fingers/hands often look wrong?

A: Hand/finger generation remains challenging for AI video models due to:

• Complex articulation
• Rapid motion during gestures
• Partial occlusion
• Perspective changes

Mitigation Strategies:

• Higher resolutions improve hand quality
• Keep hand gestures simple in prompts
• Increase sampling steps (14-16)
• Use shift=24 for better structural coherence
• Acceptance: Even commercial models struggle with hands

🔗 Related Resources

If you found this extra long generation guide helpful, explore these related tutorials:

Related Tutorials

1. LongCat Video Complete Guide Comprehensive overview of all LongCat Video features and capabilities

2. ComfyUI LongCat Workflow for Low VRAM Optimize LongCat workflows for 12GB and 16GB GPUs

3. LongCat FP8 Workflow: Memory-Efficient Video Generation Deep dive into FP8 quantization for efficient video generation

Advanced Topics

4. Painter LongVideo ComfyUI Tutorial Combine LongCat with Painter for controlled long-form generation

5. Build Your Own AI Avatar with LongCat Create consistent digital characters for extended video sequences

More ComfyUI Workflows: Visit Blog Homepage or Tutorials Category