Generate#
The Generate tab is the fastest path to an image: prompt, model and settings on the left, the viewer in the middle, the gallery on the right. Every generation is queued (see Queue) and lands in the current board of the gallery.
The launchpad#
Before the first generation the centre shows four starting points:
| Card | What it does |
|---|---|
| Generate from Text | Focuses the prompt box. The plain text-to-image path. |
| Edit Image | Drops an image onto the canvas as a raster layer for image-to-image work. See Canvas. |
| Use a Layout Image | Adds the image as a control layer so the generation follows its structure. See Control layers. |
| Add a Style Reference | Adds the image as a global reference image (IP-Adapter). See Reference images. |
Prompts#
The positive prompt describes the image; the negative prompt lists what to avoid and is only honoured by models that use classifier-free guidance (SD 1.5, SDXL, SD 3.5). Three syntaxes work inside the boxes:
- Weighting (Compel):
(golden light)1.3strengthens,(text)0.7weakens;+and-after a word do the same in steps. Available on SD and SDXL models, ignored by Flux-family models, which read natural sentences instead. - Embeddings: a textual inversion installed for the model is used by its
trigger, for example
<easynegative>in the negative prompt. - Dynamic prompts:
a {red|blue|green} carexpands to three prompts. See Dynamic prompts.
Prompt templates (the template picker above the prompt) wrap your text in a
saved style: a template contains {prompt} where your words go. Built-in
templates ship with the app; save your own with Create Prompt Template.
Structured prompts. For Flux 2 and Krea 2 models the prompt can be a JSON object built with the Prompt Builder: subjects, camera, lighting and palette as fields. Classic SD models read a JSON prompt as plain text.
Model#
The model list shows every main model installed under your models roots. Changing the model applies its recorded default settings (steps, CFG or guidance, scheduler, size) where the model carries them. The families you can install, and how they differ:
| Family | Native size | Steps and guidance | Notes |
|---|---|---|---|
| SD 1.5 | 512 | 20-30 steps, CFG 6-8 | Fast, huge choice of fine-tunes and LoRAs. Negative prompt matters. |
| SDXL | 1024 | 20-30 steps, CFG 5-7; turbo variants 4-8 steps, CFG 1-2 | Optional refiner. The upscaling tab supports SD 1.5 and SDXL only. |
| FLUX.1 dev, schnell, Krea dev | 1024 | dev: 20-30 steps, guidance 3-4; schnell: 4 steps | Reads natural language, no negative prompt. FLUX.1 dev is licensed for non-commercial use. |
| FLUX.2 Klein 4B, 9B | 1024 | 4-8 steps, guidance low | Distilled, fast, strong prompt following with the structured prompt. Klein 9B is non-commercial. |
| Z-Image Turbo | 1024 | 8 steps, CFG 1 | Very fast. Similar seeds can look alike: see Seed Variance Enhancer. |
| Krea 2 Turbo | 1024 and above | Turbo: 8 steps, CFG 1. Raw variants: 20-28 steps, CFG 3-5 | Natural-language prompts, photographic look, also the engine behind the Edit tab. Bring your own weights; see Models. |
| SD 3.5 medium, large | 1024 | 20-30 steps, CFG 4-5 | Large models; medium fits a 16 GB card. |
Formats. The same model can exist as bf16, FP8, int8 or GGUF. FP8 and int8 halve the download and the VRAM with a negligible quality change and are the recommended choice on a 16 GB card; bf16 is the reference; GGUF runs but is slower on this stack. ComfyUI's mixed NVFP4 releases (files named fp4mixed, for example the Qwen3 text encoder of FLUX.2 Klein) also load: the weights are expanded to bf16 while loading, so they save download size, not VRAM.
Seed#
The seed fixes the starting noise. Same model, prompt, settings and seed give the same image; with Use CPU Noise on (the default) the same seed also matches across machines. Leave the seed random while exploring; recall one from the gallery with Use Seed when you want to refine a result.
Iterations#
How many images to queue with the current settings, each with its own seed. With dynamic prompts on, every expanded prompt is generated this many times.
Steps, CFG and scheduler#
- Steps: denoising passes. Distilled and turbo models need 4 to 8; classic SD and Flux dev models 20 to 30. Past the sweet spot of a model, extra steps cost time without adding detail.
- CFG scale (SD, SDXL, SD 3.5) and Guidance (Flux family): how strictly the prompt is enforced. High values over-saturate and distort; distilled models want 1. Stay near the default the model set when you selected it.
- Scheduler: the sampling algorithm. Keep the default; try a different one only for a specific look, and keep the step count in the range the model expects.
Size and aspect ratio#
Pick an aspect ratio, then a size near the native resolution of the family (512 for SD 1.5, 1024 for everything else). Width and height must be multiples of 8, and 16 for the Flux and Krea families. Far above native size, classic models repeat subjects and break anatomy: use the Upscaling tab or, for Flux models, DyPE.
High resolution fix#
SD 1.5 and SDXL only. Generates at a smaller size first, then upscales with the chosen upscale method and finishes the denoise at full size, which prevents the repeated-subject artefacts of over-large first passes.
Refiner#
SDXL models can hand the last part of denoising to the SDXL refiner. Refiner start sets the hand-over point (0.8 means the base does the first 80 percent); the refiner has its own steps, CFG, scheduler and aesthetic score targets. Leave it off unless the refiner is installed and the base is SDXL.
Advanced settings#
| Setting | Use |
|---|---|
| CLIP skip | Skips the last layers of the text encoder. Some SD fine-tunes expect 2; otherwise keep 0 or 1. |
| VAE and VAE precision | Override the decoder. Use fp16 unless you see black or blown-out images, then fp32 (the fp16-fix VAE solves this for SDXL). |
| CFG rescale multiplier | For models trained with zero-terminal SNR; 0.7 tames their over-saturation, 0 disables it. |
| Seamless tiling X and Y | Textures that tile edge to edge. |
| Use CPU noise | Seed reproducibility across GPUs; no speed cost. |
| Color compensation | SDXL image-to-image and inpainting: reduces colour shifts. |
| Optimized image-to-image | Flux models: a more gradual denoising strength scale for image-to-image. |
| Text encoder on CPU | Runs the text encoder on the CPU to free VRAM for the denoiser; slightly slower. |
DyPE for large Flux images#
Dynamic Position Extrapolation lets Flux models generate well above their 1024 training size. Auto enables it above 1536 px, 4K uses settings tuned for 4K output. Scale sets the strength (default 2.0) and Exponent how fast the effect fades over the steps: 0.5 just above native, 1.0 around 2K to 3K, 2.0 for 4K.
Seed Variance Enhancer#
Z-Image Turbo produces similar images from different seeds. The enhancer adds seed-based noise to the text embeddings so seeds diverge while staying reproducible. Strength below 0.1 is subtle, 0.5 strong; Randomize percent sets how many embedding values are touched.
LoRAs#
Add LoRAs trained for the selected base model and set a weight each. See Presets and LoRAs.
Dynamic prompts#
Curly braces make the prompt a template: {red|blue|green} car becomes three
prompts. Max prompts caps the expansion. Seed behaviour: per
iteration uses one seed for all prompts of an iteration (compare wording on
the same seed), per image gives every image its own seed. Structured JSON
prompts are never expanded.
Generate#
Generate (Ctrl+Enter) queues the batch; Ctrl+Shift+Enter puts it at the front of the queue. While a batch runs, the viewer shows progress images if that is enabled in Settings. The button badge shows a count when boards are marked for batch generation.
Three habits that save time#
- Draft with a distilled model, then switch to the model you want for the final image with the same prompt.
- Explore with iterations, refine with a seed: 4 to 8 iterations to find a composition, then recall its seed and adjust the prompt.
- Reproduce from the gallery: Use All loads every setting of an image; Remix loads everything except the seed.