ComfyUI Node
Krea Reason (expand prompt + encode)
A ComfyUI node in ShootTheSound/KreaReason with 13 inputs and 2 outputs.
Krea Reason (expand prompt + encode)
- clip
- image
- conditioning
- generated_text
◄prompt►
◄modeexpand►
◄max_new_tokens220►
◄temperature0.70►
◄seed0►
◄image_removepeople + main subject (keep the scene)►
◄custom_image_instruction►
◄image_megapixels1.0►
◄instruction►
◄top_p0.95►
◄cond_boost1.0►
CategoryShootTheSound/KreaReason
Inputs (13)
| Name | Type | Default | Description |
|---|---|---|---|
| clip | CLIP | — | |
| prompt | STRING | — | |
| mode | COMBO | expand | expand: encode the rewritten prompt (impactful). think: keep prompt, append the reasoning as context (subtle/experimental). Ignored when an image is connected |
| max_new_tokens | INT | 22016–2048 | Length cap for each generation pass |
| temperature | FLOAT | 0.700–2 | 0 = deterministic (greedy). Higher = more varied phrasing |
| seed | INT | 00–18446744073709550000 | Sampling seed (only matters when temperature > 0) |
| imageopt | IMAGE | Optional reference image — described, filtered, then combined with your prompt (3 passes). Vision works on bf16 and fp8_scaled encoders; if a build can't run vision you'll get a clear error | |
| image_removeopt | COMBO | people + main subject (keep the scene) | What to REMOVE from the image description before combining with your prompt. e.g. remove people + subject to keep only the scene/background. 'custom' uses the box below. Ignored if no image |
| custom_image_instructionopt | STRING | Used only when image_remove = 'custom'. Say what to remove (or keep), e.g. 'remove the sky and any text' | |
| image_megapixelsopt | FLOAT | 1.00.1–4 | Downscale cap for the reference image (megapixels). Lower = fewer vision tokens / faster |
| instructionopt | STRING | System instruction for text-only expansion (blank = built-in). Ignored when an image is connected (image_remove drives it) | |
| top_popt | FLOAT | 0.950–1 | — |
| cond_boostopt | FLOAT | 1.00–8 | Multiply the output conditioning (CFG-free guidance boost). 1.0 = off; ~1.5-4x often sharpens Krea 2; saturates past ~6-8x. Sweep by eye |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| conditioning | CONDITIONING | — |
| generated_text | STRING | — |