ComfyUI Node
AUN KSampler PlusV3
Progressive two-pass sampler with latent and pixel-space upscale and optional final refinement pass. Flow: Base (first pass) → Latent upscaled (second pass) → Both upscaled (pixel-upscaled decoded latent, then resample mirroring second pass) → Refined image (optional). Upscaled type returns one of: 'Both upscaled', 'Latent upscaled', 'Image upscaled', or 'No upscale'. If refine is enabled, ' Refined' is appended. Each pass outputs an image: Base (first pass decoded), Image upscaled (base image upscaled in pixel space), Latent upscaled (second pass decoded), Both upscaled (pixel-upscaled decoded latent, then resampled), Refined image (final optional refinement). Right-click → "Collapse Connections" or double-click to hide slot labels and converge connection lines.
AUN KSampler PlusV3
- vae
- model
- positive
- negative
- latent_image
- Base image
- Image upscaled
- Latent upscaled
- Both upscaled
- Refined image
- LATENT
- Upscaled type
◄seed0►
◄steps_total30►
◄steps_first12►
◄start_step_second0►
◄cfg8.0►
◄cfg_latent_upscale8.0►
◄sampler_name▾►
◄scheduler▾►
◄denoise1.00►
◄latent_upscaletrue►
◄upscaling_denoise0.61►
◄upscale_method▾►
◄ratio1.50►
◄image_upscalefalse►
◄image_upscale_method▾►
◄image_upscale_model▾►
◄image_upscale_ratio1.50►
◄image_upscale_refinefalse►
◄img_refine_steps4►
◄img_refine_denoise0.25►
◄verbosefalse►
CategoryAUN Nodes/KSampler
Inputs (26)
| Name | Type | Default | Description |
|---|---|---|---|
| vae | VAE | VAE model for encoding/decoding between pixel and latent space. | |
| model | MODEL | The diffusion model to use for progressive sampling. | |
| seed | INT | 00–18446744073709550000 | Random seed for reproducible results. Use same seed for identical outputs. |
| steps_total | INT | 302–300 | Total sampling steps across both passes. Split between first and second pass. |
| steps_first | INT | 121–300 | Steps for first pass (base generation). If latent upscale is off, this is used for the single pass. |
| start_step_second | INT | 0-1–300 | Second pass control: -1 uses denoise-fraction (see 'upscaling denoise'), 0 starts at step 0, steps_first continues after pass 1. Both-upscaled mirrors this schedule to reduce drift. |
| cfg | FLOAT | 8.00–100 | CFG scale for first pass. Controls prompt adherence in base generation. |
| cfg_latent_upscale | FLOAT | 8.00–100 | CFG scale for latent upscale pass. Can be different from base CFG. |
| sampler_name | COMBO | Sampling algorithm for both passes. DPM++ 2M Karras recommended for progressive sampling. | |
| scheduler | COMBO | Noise schedule. AYS schedulers work well with progressive sampling. | |
| positive | CONDITIONING | Positive prompt conditioning (what you want in the image). | |
| negative | CONDITIONING | Negative prompt conditioning (what you want to avoid). | |
| latent_image | LATENT | Input latent image. For txt2img, use Empty Latent Image. | |
| denoise | FLOAT | 1.000–1 | Denoising strength for first pass. 1.0 = full generation, lower for img2img. |
| latent_upscale | BOOLEAN | true | Enable latent space upscaling between passes. Core feature of progressive sampling. Disabling this will perform a single pass of sampling. |
| upscaling_denoise | FLOAT | 0.610.01–1 | Second pass denoise amount. Notes: • Used only when start_step_second = -1 (denoise-fraction mode). • Typical refinement: 0.5–0.7. • Ignored when continuing from a start step. |
| upscale_method | COMBO | Algorithm for latent upscaling. Bicubic recommended for progressive workflows. | |
| ratio | FLOAT | 1.500.01–8 | Latent upscale ratio between passes. 1.5-2.0 typical for progressive generation. |
| image_upscale | BOOLEAN | false | Enable pixel-space image upscaling. Used to construct 'Both upscaled' and as source for 'Refined image'. |
| image_upscale_method | COMBO | Algorithm for final image upscaling. Lanczos for photos, nearest for pixel art. | |
| image_upscale_model | COMBO | AI upscaling model for final enhancement. 'None' disables AI upscaling. | |
| image_upscale_ratio | FLOAT | 1.500.01–8 | Final image upscale ratio. Can be different from latent upscale ratio. |
| image_upscale_refine | BOOLEAN | false | Output 'Refined image' by re-encoding and sampling the selected source (Both/Image/Base) with the settings below. |
| img_refine_steps | INT | 41–100 | Sampling steps for the final 'Refined image' pass. |
| img_refine_denoise | FLOAT | 0.250–1 | Denoising strength for the final 'Refined image' pass. |
| verbose | BOOLEAN | false | Print detailed pass logs and timings to the console. |
Outputs (7)
| Name | Type | Description |
|---|---|---|
| Base image | IMAGE | — |
| Image upscaled | IMAGE | — |
| Latent upscaled | IMAGE | — |
| Both upscaled | IMAGE | — |
| Refined image | IMAGE | — |
| LATENT | LATENT | — |
| Upscaled type | STRING | — |