Files

T

2026-07-08 15:22:16 -05:00

21 KiB

Raw Blame History

Image Oasis Help 🌴

Welcome! Image Oasis is a one-node setup for ComfyUI. Instead of wiring up 60+ nodes, you fill in the fields, hit Generate, and you're done. This guide walks the sections top to bottom - skip around as needed.

💾 Presets

A preset is a snapshot of your settings - Model, LoRA stack, Generation, Refiner, Upscale, everything in the config. It does NOT save your prompt text, your seed, or your reference images. Those are per-run, not part of a reusable "style".

To save: type a name, hit save. To load: pick from the dropdown. To delete: select it, then trash icon. To reorder: drag the 9-dot grip on the left of any preset card.

🧠 Model

This is where you pick what actually generates the image.

Three ways to load a model

Checkpoint - one .safetensors file with everything inside (model + text encoder + VAE). Simplest.
Diffusion - a standalone model file. You'll pick CLIP and VAE separately.
GGUF - a quantized model. Smaller, lighter on VRAM. Same external-CLIP/VAE setup as Diffusion.

Tell Image Oasis what kind of model you loaded so sampling is configured correctly:

AuraFlow - AuraFlow flow-matching using models like Z-Image Turbo.
Boogu-Image 0.1 (Base / Turbo) - Boogu's 10B unified image model. Turbo is 3-4 step distilled (CFG 1.0, sgm_uniform); Base uses standard flow-matching (CFG ~4, 30-50 steps). Uses the Flux VAE and a Qwen3-VL-8B text encoder with its own boogu CLIP type - the arch default sets this for you, and picking a different CLIP type will crash the run (the model needs conditioning metadata only the boogu encoder attaches). The Edit variant is not supported.
Flux.1 / Flux.2 - Black Forest Labs flow-matching.
Krea 2 (Turbo / Raw) - Krea's 12B DiT. Turbo is 8-step distilled (CFG 1.0); Raw uses standard flow-matching (more steps, positive CFG).
Qwen-Image-Edit - Alibaba's image editor. Takes reference images.
SD1 / SD1.5 / No Patch - catch-all for SD1, SD1.5, SDXL, and anything else that doesn't need a flow-matching sampling patch. This is the correct choice for SDXL.
SD3 / SD3.5 - Stable Diffusion 3 / 3.5.

💡 If you pick the wrong arch you'll get garbage or a solid gray image - the sampler is being patched for a different math regime. When in doubt, try SD1 / SD1.5 / No Patch first since "no patch" is always safe to fall back to.

⚠️ Krea 2 + Sage attention = black images. Sage attention corrupts Krea 2's math, so as of v1.4 the node refuses to run Krea 2 while ComfyUI is launched with --use-sage-attention - you'll get a clear error instead of a wasted run and a black square. Remove the flag and restart ComfyUI. Other architectures (Flux, Qwen, SD3, SDXL, etc.) are unaffected.

CLIP / VAE

If the checkpoint bundles its own CLIP and VAE, those dropdowns disappear. If not, you pick them yourself.

CLIP slots (per architecture)

The number of CLIP dropdowns you see depends on the selected architecture:

AuraFlow, Boogu-Image, Krea 2, Qwen-Image-Edit - one CLIP slot.
Flux.1 / Flux.2, SD1 / SD1.5 / No Patch - two slots (slot 2 optional). Flux uses clip_l + t5xxl; SDXL uses clip_l + clip_g; SD1.5 / Flux.2 Klein leave slot 2 empty.
SD3 / SD3.5 - three slots (clip_l + clip_g + t5xxl). The triple-CLIP combo produces noticeably better quality than dual on SD3.5.

Empty slots are ignored - leave them blank if your config doesn't need them. Your CLIP picks persist across architecture switches: jump from SD3 to Qwen and back and your three SD3 CLIPs are still there.

💡 GGUF + safetensors can mix. A GGUF CLIP in slot 1 and a safetensors CLIP in slot 2 (or any combo) works fine - the GGUF loader handles the fusion.

🎨 LoRA

LoRAs are little add-ons that nudge the model toward a particular style, character, or concept.

+ Add LoRA drops a new row.
Strength - typical range 0.5-1.5. Depends on LoRA. Consult Civitai page for recommended LoRA settings.
Toggle - enables/disables without removing.
Trigger words - some LoRAs need a specific word or phrase in the prompt to activate. Type it in the trigger field under each enabled LoRA and it's automatically prepended to your positive prompt at run time. Leave blank for LoRAs that don't need a trigger. Disabled LoRAs are skipped.
Drag handle (9-dot grip) - reorder. LoRAs apply top to bottom, so the order matters when two LoRAs compete.
✕ - delete the row.

💡 If two LoRAs fight, try reordering before lowering strengths.

💡 If a LoRA isn't doing anything visible, check whether it needs a trigger word. Trigger-word LoRAs are silent without one.

🖼️ Reference Images

Two groups of slots, used by two different features:

Qwen Image Edit (slots 1-3)

Feed the Qwen-Image-Edit architecture's edit conditioning - "take this picture and change X". Up to three slots; most edit models only use one. The Fit Method in the Latent section controls how these are mapped onto your output size. On any other architecture these slots dim and are ignored.

Img2Img Init (non-Qwen)

One Init slot. Fill it on any non-Qwen architecture (Flux, SD1.5/SDXL, SD3, AuraFlow, Krea 2) and generation starts from this image instead of from pure noise. Occupied slot = img2img on; clear the slot to return to normal generation. There is no separate toggle.

What img2img actually is: think image inspiration, not image editing. Any diffusion model can do it - the model remakes your image in its own signature style, carrying over the structure, composition, and palette. Every model's style bleeds onto the result (a Flux remake looks like Flux, a Krea remake looks like Krea), and models differ in how strongly they impose it versus preserve the original. For surgical "change only the hat" instructions, that's Qwen-Image-Edit's job.

Strength is the Denoise value in the Generation section. Start at 0.5 and adjust from there - the right value depends entirely on what you're asking the model to do to the image. Lower preserves more of the source; higher hands more over to the prompt and the model's own style.
Sizing is the Fit Method in the Latent section.
On the Qwen arch this slot dims and is ignored - Qwen already consumes images through the edit slots above.

Three ways to fill a slot

Click the upload button.
Drag and drop a file onto the slot's image box - from your file manager, from the ComfyUI Asset side panel, or even from the node's own output pane (dragging your last output into Init is a quick way to iterate on it).
Paste - click the slot's image box, then Ctrl+V. Works with images copied straight off a website or from a screenshot tool.

💡 The little square thumbnail is the drop & paste target - it lights up when it's ready to receive. Hover it for a reminder.

Image info line

Occupied slots show the image's resolution and file size under the upload button. The ⤢ button next to it copies the image's dimensions into the Latent Width/Height (snapped to /16) - one click to make the output the same size as the source.

Other slot controls

✕ - clears the slot.
◧ - uses this image as the Compare base in the output viewer (see Output section below). Works on all four slots, including Init - handy for wiping between your source and the img2img result. Only one slot can be the compare base at a time.

✍️ Prompt

User Prompt

Your short, sketch-style prompt - the source the enhancer rewrites from. This box is sticky: clicking 🪄 Enhance doesn't change it, so you can re-click Enhance over and over to iterate against the same starting point without losing your original.

💡 If you're not using the enhancer, you can leave User Prompt empty and just type your full prompt into Enhanced Prompt below.

Enhanced Prompt

The expanded version of your User Prompt - this is what actually drives image generation. Two ways it gets populated:

Click 🪄 Enhance and the LLM writes here.
Type or paste directly. Hand-edits stick around until the next Enhance click overwrites them.

Every Enhance click overwrites this box with a fresh result. That's the designed iteration loop - re-click for a new variation, or hand-tweak the result.

Negative prompt

What you don't want. Only used when CFG > 1. At CFG = 1 the box grays out and gets ignored - there's no path for it.

Prompt Enhancer 🪄 (magic wand)

A local LLM that rewrites your User Prompt into a detailed Enhanced Prompt. Runs out-of-band: it is not part of the generation graph. The model stays loaded between enhance clicks (so re-clicking is fast) and unloads automatically when you queue an image, so it never fights the diffusion model for VRAM during generation.

Pick an enhancer model from the dropdown. (Drop .gguf files in ComfyUI/models/LLM/ to populate it.)
Pick a style: Natural Language (a flowing paragraph, best for Flux / Qwen-Image-Edit / SD3) or Tags (comma-separated, best for SDXL-style models).
Type a short prompt into User Prompt.
Click the wand 🪄. The result lands in Enhanced Prompt.
Want a different take? Click the wand again - the LLM re-enhances from your User Prompt and overwrites the Enhanced Prompt with a fresh result.

💡 The first Enhance click is slower (the model loads). Subsequent clicks reuse the loaded model and are fast. The model only unloads when you queue an image generation.

Auto cleanup

Different LLMs leak different artifacts in their output - a <think>...</think> reasoning block, an orphan </think> tag, stray [INST] / [OUT] / <<SYS>> template tokens. Image Oasis strips these for you automatically: there's a profiles.json next to routes_enhance.py with a list of cleanup profiles, each matching a model filename pattern. First profile that matches wins; nothing matches, a universal rule set runs. Edits to profiles.json take effect on the next Enhance click - no ComfyUI restart needed.

💡 If a new model is leaking junk in the output, add a profile for it in profiles.json and click Enhance again. You don't need to touch any code.

Enhancer Settings (collapsible sub-panel at the bottom of the Prompt section)

Click the Enhancer Settings header to expand. Three knobs, all optional - defaults are tuned for the curated model set.

Auto GPU layers (checkbox, on by default). When on, Image Oasis reads the GGUF header for the model's layer count, checks live free VRAM, and picks how many layers to put on the GPU - shown next to the checkbox as Recommended: All (N) when the whole model fits, or Recommended: K/N when only a partial offload fits. The reading happens after evicting the diffusion model, so it reflects what will actually be free at enhance time.
GPU layers (number field). Manual override - only editable when Auto is off. -1 = put everything on the GPU and let llama.cpp figure it out. Anything 0 or higher is a literal layer count.
Context (number field). The LLM's context window size (n_ctx). Default 8192. Bigger = longer inputs/outputs work, but uses more VRAM at load time.
Max tokens (number field). The maximum number of tokens the LLM will emit per Enhance click. Default 2048. Capped internally against what your Context can actually hold after the input, so over-large values just clamp.

⚠ Only disable Auto GPU layers if you know what you're doing. An over-aggressive manual value will OOM the model load. Image Oasis retries on CPU as a backstop, but you'll have given up the GPU speedup.

Picking an enhancer model

Use an abliterated GGUF - at the largest size + quant that fully fits your VRAM.

Recommended models for low-VRAM setups:

Model                            Quant  Tier
Llama3.3 8B Instruct Abliterated Q5_K_M (A)
Mistral 7B Instruct Abliterated  Q5_K_M (S)
Qwen2.5 7B Instruct Abliterated  Q5_K_M (C)
Qwen3 8B Abliterated             Q4_K_M (C)
Qwen3.5 9B Abliterated           Q4_K_M (B)
SuperGemma4 E4B Abliterated      Q5_K_M (A)
WizardLM-2 7B Abliterated        Q5_K_M (S)

📐 Latent

The canvas: what size and shape the generation starts as.

Width / Height

Output dimensions in pixels. The ↔ button swaps them. Every image that enters generation - the img2img init or the Qwen edit references - is conformed to this size, so what these fields say is what you get.

Ratio presets

Toggle a ratio button (1:1, 2:3, 3:4, 9:16, 16:9, 4:3, 3:2) to lock the aspect: the height snaps immediately, and from then on editing either field recalculates the other (to the nearest multiple of 16 - the safe snap for every supported architecture). Click the active button again to unlock and edit freely. The ↔ swap also flips the lock to its mirror (2:3 becomes 3:2) so the highlighted button stays honest.

Fit Method

How incoming images are conformed to your Width x Height. On the Qwen-Image-Edit architecture this applies to the edit reference images; on every other architecture it applies to the img2img Init image. When an image path is in effect, a line under the buttons says which one - the Qwen edit references on Qwen, the img2img init otherwise.

Stretch - scale to exactly the target size. Ignores aspect; a mismatched image distorts.
Crop (default) - scale to fill preserving aspect, then center-crop the overflow. No distortion; the edges may be lost.
Pad - scale to fit inside preserving aspect; the leftover space is filled by repeating the image's own edges. The padded regions are invented by the model, so expect the borders to be made up.

💡 If you want the output shaped like your source image, click the ⤢ button on the image's info line (see Reference Images) or the matching ratio preset - then every Fit Method behaves identically because nothing needs conforming.

Batch

Generate multiple images in one run. Each gets a different starting noise pattern (seed).

⚙️ Sampling

The actual image-generation step.

Steps

How many denoising iterations. More = more refined, with diminishing returns. Use recommended model settings.

CFG (Classifier-Free Guidance)

How strictly the model follows the prompt. Higher = stricter, lower = looser. Use recommended model settings.

💡 CFG = 1 is a special fast mode. Only one forward pass per step (vs. two), so it's roughly 2× faster. Distilled/turbo models are trained for this. Negative prompt is disabled at CFG = 1.

Sampler / Scheduler

The math used to walk through noise removal. Some common pairs:

euler + simple - safe default. fast iteration
dpmpp_2m + karras - quality favorite. slower iteration.
lcm / sgm_uniform - for distilled models.
euler_a / beta - my personal favorite.

Results vary by model - experiment.

Denoise

How much noise to start from. 1.0 = fresh generation. Lower values keep more of the starting latent - this is the strength dial for img2img (Init slot occupied): start at 0.5 and adjust from there, since the right value depends entirely on what you're asking the model to do to the image. Also used internally by the refiner pass.

Variety

Fixes the "every seed looks the same" problem on distilled models (Z-Image Turbo, Krea 2 Turbo, Boogu Turbo, and similar). On these models the prompt dominates so strongly that re-rolling the seed barely changes the composition. Variety adds a tiny, seeded amount of noise to the prompt conditioning during the early sampling steps - just enough to land each seed on a genuinely different layout - and hands back the clean prompt for the rest of the run, so adherence and fine detail are unaffected.

0 = off (default; exact current behavior).
Start around 0.1 and adjust. Higher values diversify harder but drift further from the prompt.
Fully reproducible: the same seed with the same Variety value regenerates the same image.

Shift

A flow-matching parameter. 0 = use the architecture's recommended default (recommended unless you know the model). Hover the field for typical ranges by architecture.

Seed

The random number that locks in the noise pattern. Same seed + same settings = same image.

🎲 - randomize and run.
▶ - run with the current seed (re-iterate without changing the noise base, helpful when refining or upscaling a generated image).

After-generate

What happens to the seed after each run:

Fixed - stays the same.
Increment / Decrement - steps by 1.
Randomize - new seed every time.

🔬 Refiner

A second sampling pass that runs on top of the first one.

When to use it

Iterating on an existing image - enable refiner with a low denoise (~0.1-0.3) to nudge the current image rather than fully regenerate.

The refiner has its own steps, CFG, and denoise - all independent from the base pass.

🔍 Upscale

Make the output image bigger after generation.

Two modes

Algorithmic - fast classic resizing (bilinear, bicubic, lanczos, nearest, bislerp). No model needed.
Model - a dedicated upscale model (ESRGAN-family, e.g. 4x-UltraSharp). Slightly slower but much better quality, especially on real-world detail.

×Multiplier

How much to scale. 2× doubles resolution; 4× quadruples. With model mode, the model's "native" scale is applied first, then resized to your target multiplier if it doesn't match. This way, you can use a x4 upscale model but still only upscale the image to whatever size you want.

💡 OOM-safe. Image Oasis tiles automatically and falls back to smaller tiles if it runs out of VRAM mid-upscale.

🎨 Theme

Customize the node's colors. Six theme variables drive everything:

Accent - the bright highlight color (R&G button fill, active toggles, etc.).
Accent (dim) - darker variant of Accent (R&G/Save button fill, dim states).
Background - the node's main background. Also drives the Generate/Compare button fill.
Panel - section headers.
Border - borders everywhere. Also drives the Generate/Compare button border.
Muted text - labels, hints, dim text.

Named themes

Save a palette under a name. Switch between them with one click. Themes apply to every Image Oasis node in the workflow at once.

Reset

The Reset to Default button clears all overrides - back to the default slate-blue + dark palette.

📤 Output (the right side)

Where generated images appear.

Header buttons

▶ Generate - runs with the current seed. Useful when adding a refiner/upscale pass to the last generated image.
🎲 Randomize & Generate - picks a new seed and generates a new image.
⏹ Interrupt - stops the current run. Only appears while a generation is in progress.
◧ Compare - toggles the A/B slider (see below).
💾 Save - copies to your ComfyUI output folder.

Execution timer

The big MM:SS:mmm readout glows while a run is in progress; freezes at the final time when done. It survives tab switches - even mid-generation: switch away while a run is live and the clock picks back up from the true elapsed time when you return.

When a batch returns more than one image, ‹ N/M › appears in the bottom-right of the footer. Click the arrows to step through them. Wraps around at both ends. The compare slider (if active) pairs by index - image N current vs. image N previous.

Compare slider (the ◧ button)

Drag the white handle to wipe between two images.

Default: current image vs. the previous generation. Toggle refiner or upscale on, hit ▶ Generate, then ◧ - you can drag the slider bar to preview exactly what changed.
Versus a reference image: click the ◧ next to a reference slot. The slider now compares current vs. that reference image. Click again to clear.
Batches: every image in the batch gets its own slider (paired by index) when the previous batch had the same count. They all share one slider position so you can compare the whole batch at once.
Hidden: if there's no previous and no ref toggled, the slider just doesn't appear.

Tips & gotchas 💡

VRAM tight? Use GGUF models - they're quantized to fit in less memory.
Generation feels slow? Try CFG = 1 with a distilled / turbo model. Roughly 2× faster.
Same image every run? Set the after-generate action to Randomize or use 🎲 Randomize & Generate.
Negative prompt seems to do nothing? Check CFG - at 1.0 it's intentionally ignored.
Strength tweaks are cheap. Changing a LoRA strength re-patches from cache - only changing the model file triggers a full reload.
Got an error instead of an image? That's on purpose. When something genuinely fails mid-run (most often running out of VRAM), the run stops and shows the error rather than handing you a gray image. Fix the cause (smaller batch, lower resolution, GGUF quant) and re-queue - failed runs are never cached, so retrying actually retries.

"Accessibility tool for the nodally challenged" - PheebyKatz 💖

21 KiB Raw Blame History Unescape Escape