Fast video in four steps

<!-- lesson: video-fast -->

Lesson "video-fast", in the group Video of the lesson list. Steps: 8.

#What you will be able to do

You cut twenty steps down to four with a speed-up LoRA.

#What you need

• It goes on from LoRA: add a style or a subject. The card says so; it is a hint, not a lock.
• Model files the lesson checks:
  • the high_noise half of Wan 2.2: models/diffusion_models, for example Wan2_2-T2V-A14B-HIGH_fp8_e4m3fn_scaled_KJ.safetensors;
  • the low_noise half of Wan 2.2: models/diffusion_models, for example Wan2_2-T2V-A14B-LOW_fp8_e4m3fn_scaled_KJ.safetensors;
  • a speed-up LoRA for Wan (lightx2v or similar), its high file: models/loras, for example wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors;
  • a speed-up LoRA for Wan (lightx2v or similar), its low file: models/loras, for example wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise.safetensors.
• Nodes: Load Diffusion Model, ModelSamplingSD3, LoraLoaderModelOnly, KSampler (Advanced).
• The app downloads and installs nothing. The step texts below say which files and folders, and which ComfyUI template names them.

#The steps

The words of each step, as the lesson shows them. Node names are in bold.

1. Where the speed comes from (reading)

   A speed-up LoRA teaches the model to do in four steps what usually takes twenty. The best known one for Wan is called lightx2v.

   The price: less variety and slightly rougher motion. For trying things out and finding the right prompt it nearly always pays: the clip is ready several times sooner.

2. A video graph is needed

   This lesson goes on from the graph of "Video from text" or "Video from a picture". The canvas must have two Load Diffusion Model nodes and two ModelSamplingSD3 nodes.

   If they are not there, do one of those lessons first.

3. Put a LoRA into both models

   Two LoraLoaderModelOnly nodes are needed, each between its Load Diffusion Model node and its ModelSamplingSD3 node. How to put them in is shown in "LoRA: add a style or a subject".

4. Choose the speed-up files

   In the node in the path of the high_noise model choose the lightx2v file with high in its name in lora_name, and in the path of low_noise the file with low.

   The files for t2v and i2v models differ: take the ones that fit your models. Leave strength_model at 1.

5. The two samplers

   The two samplers of a video lesson are needed: the first takes the high_noise model, the second takes the low_noise model and the result of the first.

6. Set four steps

   Set steps to 4 in both samplers. Set end_at_step to 2 in the first and start_at_step to 2 in the second.

   Four steps are split in half: the first model does two, the second does two.

7. Set cfg to 1

   Set cfg to 1 in both samplers.

   A speed-up LoRA is trained to work without text guidance. With a larger value the clip burns out and fills with artifacts. At cfg 1 the text of the negative input stops acting.

8. Now it is fast (reading)

   The clip is made several times sooner.

   Once the prompt is right you can go back to twenty steps for quality: take the speed-up LoRAs out and set steps back to 20, cfg to 3.5 and the hand-over to step 10.

#What the lesson checks

A reading step is done when you press Got it, next. A doing step ticks itself off when the canvas holds:

• Step 2, A video graph is needed:
  • a Load Diffusion Model node.
  • a Load Diffusion Model node.
  • a ModelSamplingSD3 node.
  • a ModelSamplingSD3 node.
  • unet_name of Load Diffusion Model (first pass) holds the high_noise half of Wan 2.2; such a file lies in models/diffusion_models, for example Wan2_2-T2V-A14B-HIGH_fp8_e4m3fn_scaled_KJ.safetensors.
  • unet_name of Load Diffusion Model (second pass) holds the low_noise half of Wan 2.2; such a file lies in models/diffusion_models, for example Wan2_2-T2V-A14B-LOW_fp8_e4m3fn_scaled_KJ.safetensors.
• Step 3, Put a LoRA into both models:
  • a LoraLoaderModelOnly node.
  • a LoraLoaderModelOnly node.
  • MODEL of Load Diffusion Model (first pass) into model of LoraLoaderModelOnly (LoRA of the first pass) (a LoraLoaderModelOnly or ModelSamplingSD3 in between is accepted).
  • MODEL of LoraLoaderModelOnly (LoRA of the first pass) into model of ModelSamplingSD3 (shift of the first pass) (a LoraLoaderModelOnly in between is accepted); or MODEL of ModelSamplingSD3 (shift of the first pass) into model of LoraLoaderModelOnly (LoRA of the first pass) (a LoraLoaderModelOnly in between is accepted).
  • MODEL of Load Diffusion Model (second pass) into model of LoraLoaderModelOnly (LoRA of the second pass) (a LoraLoaderModelOnly or ModelSamplingSD3 in between is accepted).
  • MODEL of LoraLoaderModelOnly (LoRA of the second pass) into model of ModelSamplingSD3 (shift of the second pass) (a LoraLoaderModelOnly in between is accepted); or MODEL of ModelSamplingSD3 (shift of the second pass) into model of LoraLoaderModelOnly (LoRA of the second pass) (a LoraLoaderModelOnly in between is accepted).
• Step 4, Choose the speed-up files:
  • lora_name of LoraLoaderModelOnly (LoRA of the first pass) holds a speed-up LoRA for Wan (lightx2v or similar), its high file; such a file lies in models/loras, for example wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors.
  • lora_name of LoraLoaderModelOnly (LoRA of the second pass) holds a speed-up LoRA for Wan (lightx2v or similar), its low file; such a file lies in models/loras, for example wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise.safetensors.
• Step 5, The two samplers:
  • a KSampler (Advanced) node.
  • a KSampler (Advanced) node.
  • MODEL of ModelSamplingSD3 (shift of the first pass) into model of KSampler (Advanced) (first pass) (a LoraLoaderModelOnly in between is accepted).
  • MODEL of ModelSamplingSD3 (shift of the second pass) into model of KSampler (Advanced) (second pass) (a LoraLoaderModelOnly in between is accepted).
  • LATENT of KSampler (Advanced) (first pass) into latent_image of KSampler (Advanced) (second pass).
• Step 6, Set four steps:
  • steps of KSampler (Advanced) (first pass) is 4. Why: a speed-up LoRA is trained to do everything in four steps.
  • steps of KSampler (Advanced) (second pass) is 4. Why: a speed-up LoRA is trained to do everything in four steps.
  • end_at_step of KSampler (Advanced) (first pass) is 2. Why: four steps are split in half: the first model does two, the second does two.
  • start_at_step of KSampler (Advanced) (second pass) is 2. Why: four steps are split in half: the first model does two, the second does two.
• Step 7, Set cfg to 1:
  • cfg of KSampler (Advanced) (first pass) is 1. Why: a speed-up LoRA is trained without prompt guidance. With a higher cfg the clip burns out.
  • cfg of KSampler (Advanced) (second pass) is 1. Why: a speed-up LoRA is trained without prompt guidance. With a higher cfg the clip burns out.

#Mistakes people make

• cfg left at 3.5: burnt, noisy clip.
• t2v speed-up files with i2v models, or the other way round.
• Leaving 20 steps: no speed gain, often over-cooked results.
• Counting on the negative prompt: at cfg 1 it does nothing.

#See also

• How lessons work
• Speed-up LoRAs
• CFG and guidance
• Steps
• Wan 2.2 two passes