AI videos might look impressive for the first few seconds, but as the video gets longer, things start to fall apart. Characters subtly change faces, objects lose shape, motions become jittery, and eventually the whole video feels off. Most video models are trained on perfect, clean data. But during real generation, they have to rely on their own previous frames. Once a small mistake sneaks in, the model treats it as truth and the next frame gets even worse. Over long videos, this becomes visual chaos. Stable Video Infinity removes the problem using a training strategy called Error Recycling Fine Tuning.
 |
| Stable video infinity architecture (Ref-research paper) |
The researchers let the model generate frames with errors. Then stored those imperfect outputs in a buffer, feed them back during training as conditioning inputs and teach the model how to recover from its own mistakes. This works by making error correction part of the learning process itself. And the result was strong temporal consistency, reduced error accumulation with videos that do not degrade, even as they get longer. You can find more detailed insights by accessing their research paper.
This approach doesnot increase inference cost. Once trained, the model runs like a normal video generator but behaves far more robustly over time. The system also works across different conditioning types like text prompts, audio, or motion inputs, making it flexible for real world applications.
Installation
1. Update ComfyUI to its latest version from the Manager.
2. Make sure you have already installed the Kijais' ComfyUI WanVideoWrapper custom node and have the basic Wan 2.2 I2V High and Low noise, Lightx2v Lora, Vae, text encoders etc models already downloaded and setup. If already done, then update this custom node to its latest version from the Manager by selecting Custom nodes manager option.
3. Now, download SVI Model V2 Pro. This fixes video glitches, weird artifacts, color degradation etc.
(a) SVI_Wan2.2-I2V-A14B_high_noise_lora_v2.0_pro.safetensors
(b) SVI_Wan2.2-I2V-A14B_low_noise_lora_v2.0_pro.safetensors
Choose the High and Low V2 Pro variant, as it removes the artifacts with repetitive effects over the older variants. Save them into ComfyUI/models/loras folder.
Alternative:
You can also use Kijai's optimized Wan 2.2 SVI Lora v2.0:
(a) SVI_v2_PRO_Wan2.2-I2V-A14B_HIGH_lora_rank_128_fp16.safetensors
(b) SVI_v2_PRO_Wan2.2-I2V-A14B_LOW_lora_rank_128_fp16.safetensors
Save them into ComfyUI/models/loras folder.
4. Restart and refresh ComfyUI.
Workflow
1. Download the basic 10/5 second workflows from official github repository.
2. Drag and drop into comfyui. Load the workflow and install the missing custom nodes from the manager by selecting Install Missing custom nodes.
3. Load your different input images into different Load image nodes. Load the models into their respective nodes.
4. Add the prompts for longer video generation in text prompt streaming techniques (for multiple batches). Means add text prompts for first video clip, then add second text prompt for second and so on.
The workflow is designed for infinite video generation. You can add copy and add multiple groups to make your video further longer.
Settings-
CFG: 1.5
Shift: 8
Steps: 6
Sampler: Euler
Scheduler: Simple
Based on the Stable Video Infinity research paper, prompting is not just about writing a good text prompt, it's about how you structure continuity, transitions, and intent over time. Below is a practical, user-focused guide on how to prompt for best results-
1. Think in Segments, Not One Giant PromptEx- A cinematic video of a man walking through a city, then a forest, then a beach, then a sunset
2. Keep the Core Identity Constant Across PromptsA middle-aged man with short black hair, wearing a gray jacket.Donot suddenly switch to:A young man with dark hair in a hoodie
3. Use Temporal Language ExplicitlyThe camera slowly moves forward as the lighting gradually becomes warmer.
4. Describe Motion and Change, not Just Appearance.
Instead of- A realistic forest scene. Try-A realistic forest where leaves gently sway, light shifts as clouds pass, and the camera subtly pans forward.
5. Let the Model Correct Itself (Donot Over Constrain)-Avoid excessive negative prompts,Overly strict camera instructions every frame or repeat 'perfect, flawless, ultra-perfect' constantly.
6. For Transitions, Describe Cause, Not Just Outcome. Instead of-The scene changes from city to forest. Try-As the man walks forward, buildings slowly fade into trees, traffic noise softens into wind and birds.
7. Use Stable Style Anchors. If you want consistent aesthetics across long videos, lock the style early and donot keep changing it.
8. Audio / Motion / Skeleton Conditioning- Keep Prompts Aligned if you are using audio prompts, motion inputs Skeleton guidance. Make sure your text prompt matches them semantically.
