ComfyUI Node
Qwen Rectified Flow Inverter
A ComfyUI node in Qwen/Sampling with 10 inputs and 1 output.
Qwen Rectified Flow Inverter
- model
- latent_image
- positive
- LATENT
◄seed0►
◄steps20►
◄inversion_strength0.50►
◄velocity_amplification0.00►
◄velocity_perturb_strength0.00►
◄internal_precisionbfloat16►
◄normalize_outputenable►
CategoryQwen/Sampling
Inputs (10)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | UNet to integrate during the rectified flow inversion. | |
| latent_image | LATENT | Latent to be inverted toward the clean state. | |
| positive | CONDITIONING | Positive prompt conditioning applied during inversion. | |
| seed | INT | 00–18446744073709550000 | Seed for the random velocity perturbation. |
| steps | INT | 201–1000 | Number of integration steps available for the inverter. |
| inversion_strength | FLOAT | 0.500–1 | Fraction of the schedule to walk backward during inversion. |
| velocity_amplification | FLOAT | 0.00-0.9–2 | Deterministically amplifies (>0) or dampens (<0) the predicted velocity at each step. |
| velocity_perturb_strength | FLOAT | 0.000–5 | Adds seeded random noise to the velocity. >1.0 will heavily randomize the path. |
| internal_precision | COMBO | bfloat16 | Autocast precision used while running the inverter UNet. |
| normalize_output | COMBO | enable | Crucial for high inversion strengths. Rescales the output latent to a standard distribution (mean=0, std=1) to prevent the sampler from collapsing. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| LATENT | LATENT | — |