Krea2_ComfyUI_Advanced Introduction
Krea2_ComfyUI_Advanced is a collection of ComfyUI nodes for creating and editing images with Krea 2. It brings text-to-image generation, reference-image prompting, image-to-image editing, and high-resolution workflows into one node pack. You can begin with a short prompt and a fast generation, then add more control only when your artwork needs it.
The extension is designed around Krea 2’s genuine Hugging Face Krea2Pipeline. It also includes its own sampling tools, so its workflows do not depend on another sampler node pack. Its main uses include:
- Creating images from text, with prompt-building tools and optional local prompt expansion.
- Refining images at larger sizes, using up to three generation stages or a 2× upscale decoder.
- Working from existing images, whether you want to preserve a composition, suggest a style, guide depth or pose, repaint an area, or make an instruction-based edit.
- Comparing creative choices, such as samplers, prompts, LoRAs, and checkpoints, in organized visual sweeps.
- Reusing successful settings through presets and generation settings saved with an image.
The author provides a multi-purpose example workflow. Drag its PNG onto the ComfyUI canvas to load the embedded graph. Its groups can be bypassed with Ctrl+B, letting you try different tasks without rebuilding the workflow.
How Krea2_ComfyUI_Advanced Works
Think of generation as developing an image from a rough sketch. Krea 2 begins with a hidden image representation called a latent, then gradually turns noise into an image guided by your prompt. A sampler determines how it moves through those steps; a schedule determines how those steps are distributed. Finally, a VAE decoder turns the latent into a viewable image.
The extension gives you several ways to shape that process:
- Describe the image. Write a natural-language prompt, or use Eric Krea2 Prompt to assemble descriptions of subjects, materials, lighting, and camera choices. Krea 2 works with descriptive language rather than JSON-style instructions.
- Choose a starting point. Generate from noise for text-to-image, or use Eric Krea2 VAE Encode to turn an existing image into an
init_latentfor image-to-image work. - Choose how much to develop or change it. Eric Krea2 Multi-Stage Ultra can run up to three stages. For example, Stage 1 can establish the composition, while a later stage refines a larger version. In image-to-image work, the stage’s re-noise step window determines how much room the model has to change the source.
- Add optional guidance. A reference picture can influence the subject or visual ideas through Vision Prompt. Other nodes can guide depth or pose, supply style, mark an area to repaint, or provide an editing instruction.
- Decode and save. Choose a standard decode or the optional 2× upscale decoder. Settings can be embedded in the saved PNG so you can revisit the recipe later.
Two reference-image methods are especially important to distinguish:
- Vision Prompt lets Krea 2’s Qwen3-VL text encoder interpret a reference picture. It is useful for visual ideas such as “a ceramic vase with this shape and color palette,” but it does not preserve the reference’s exact pixels.
- Reference Latents (Edit) passes encoded reference-image information into an edit pathway for stronger image or structural conditioning. It requires an edit-trained LoRA. Using it with only the base model can produce a blocky mosaic rather than a useful edit.
For a closer look at these workflows, see Image-to-image and reference guidance.
Krea2_ComfyUI_Advanced Features
Generation and high-resolution workflows
- Eric Krea2 Generate is the straightforward text-to-image node. Its automatic settings select 8 steps and guidance 0 for Turbo, or 28 steps and guidance 4.5 for Raw.
- Eric Krea2 Multi-Stage Ultra (res) offers up to three progressive stages. You can choose each stage’s sampler, schedule, guidance, and re-noise window. Its resolution-chain readout shows the actual sizes through S1 → S2 → S3 → decode, helping you see what will be generated rather than assuming every stage has the final size.
- Eric Krea2 Multi-Stage Ultra V2 (presets) saves the node’s settings—including sigma and size settings—as one reusable recipe.
- Eric Krea2 Resolution selects a size from an aspect ratio and megapixel target, or from a source image. Latent Resize changes latent dimensions without decoding and re-encoding.
There are two principal ways to reach a larger image. Turbo with 2× Upscale Decode is the fast path: the decoder enlarges the result without another denoising pass. Raw with multi-stage re-denoising gives you more control over how the model develops the larger image. The upscale VAE can also decode at 2× and downsample back to the original size as a quality pass rather than a size increase. See High-resolution strategies.
Samplers, schedules, and comparison tools
Ultra lets you select a sampler per stage: euler, res_2m, res_2s, deis_3m, abnorsett_3m, abnorsett_4m, lawson4_4s, etdrk4_4s, er_sde, rk6_7s, lcm, lcm_hybrid, or lcm_hybrid2. Available schedules include linear, balanced, karras, beta57, beta, bong_tangent, exponential, linear_quadratic, and simple. Eric Krea2 Sigmas adds per-stage curve and detail-bias controls, with a live preview of where sampling and re-noising occur.
You do not need to master every option at once. Start with an existing recipe, change one setting, and compare the images. For broader comparisons, Sweep Plan can test grids or hand-written combinations of settings and produce contact sheets, per-image PNGs, and a report. Sweep Checkpoints compares multiple Turbo checkpoints in the same sweep; Sweep → Preset turns a chosen result into a reusable Ultra preset. See Samplers and schedules and the sweep guide.
Reference images and image editing
The image-to-image builder nodes connect to sockets on Multi-Stage Ultra:
| Tool | What you can use it for | Useful adjustment |
|---|---|---|
| Vision Prompt | Let one to three pictures inform the prompt’s visual meaning. | Combine it with an init_latent when you want the starting image to supply structure and the reference to suggest content or style. |
| Control (Depth/Pose ControlNet-LoRA) | Guide the output with depth or pose information using suitable Control-LoRA weights. | Set different strengths by stage—for example, stronger composition control in S1 and less control during final detail. |
| Style Reference | Borrow visual style from an image without training a style LoRA. | Use its recommended or custom preset and stage multipliers; its default emphasis is on S1. |
| Paint | Inpaint a masked area or extend a canvas with outpainting. | Choose the adapter and per-stage restore settings; inspect its canvas preview and generated mask. Paint Composite can paste the original pixels back outside the mask with a feathered edge. |
| Edit | Apply an instruction to a source image, optionally with a second reference, using the Krea 2 Identity Edit pathway. | Adjust fitting and reference-grounding controls when the edit needs a different balance of change and preservation. |
| Reference Latents (Edit) | Supply encoded reference images to an edit-trained LoRA. | Use this only with a compatible edit-trained LoRA. |
Not every guidance method can run together. Control can accompany style, paint, edit, or reference latents; style, paint, edit, and reference latents have restrictions when paired with one another. Negative guidance is skipped with style, paint, or reference latents. If a combination is unsupported, Ultra reports an error or notes that guidance was skipped. Check image-to-image guidance before building a complex combination.
Prompts, negative guidance, and LoRAs
Eric Krea2 Prompt combines structured descriptions into natural language and quotes text you want rendered in the image. Magic Prompt can expand a brief idea using a local LLM endpoint. For best compatibility with Krea’s prompt expansion, the author recommends Krea’s expansion.txt as its system prompt.
For Turbo, conventional negative guidance is not the only option. Eric Krea2 Negative Guidance provides NAG and NegPiP. NegPiP can use negative phrases in the form (phrase:-1) or its negpip_negative field; NAG has its own negative-text and strength controls. They can discourage unwanted features even when Turbo’s guidance is 0. See Negative guidance for tuning advice.
LoRAs are optional add-ons that steer a model toward a learned look or capability. Apply LoRA handles one adapter, while Multi-LoRA Stack combines several. Both support different weights for S1, S2, and S3: a style might have a stronger influence while the composition forms, then a lighter influence during detail refinement. You can apply, unload, or diagnose adapters, save named presets, search LoRA lists, and extract <lora:name:weight> tags from prompts with Prompt LoRA Tags. See LoRA workflows.
Decoding, presets, and saved recipes
VAE Decode produces a standard 1× image; Upscale Decode (2×) uses the optional spacepxl Wan 2× upscale VAE. DeGrid can remove a visible 2-pixel VAE lattice and show a report of what it removed. Both decode nodes also offer a degrid toggle.
Named presets are available for settings such as loaders, LoRAs, Ultra, and sigmas. Merge Settings gathers recipe sections into settings that can be embedded in a saved PNG; Settings from Image reads those settings back. This makes it easier to return to an image’s generation choices instead of relying on memory.
Settings merged for a metadata-aware image save
Krea2_ComfyUI_Advanced Models
The two main Krea 2 checkpoint choices are:
| Model | When to choose it | What to expect |
|---|---|---|
| Krea 2 Turbo | Fast image generation and the 2× upscale-decode path. | The author’s automatic generation settings use 8 steps and guidance 0. It can produce glossy or over-pronounced fine texture in some images, especially close-up skin. |
| Krea 2 Raw | More controlled, multi-stage re-denoising and LoRA training. | This undistilled base is suited to slower refinement; automatic generation uses 28 steps and guidance 4.5. The author’s guidance is to train LoRAs on Raw and run them on Turbo. |
Both use Krea 2’s image model and Qwen3-VL conditioning. The Component Loader also supports compatible component files and checkpoint formats, including bf16, GGUF, fp8, INT8, MXFP8, and NVFP4 variants. Quantized checkpoints can reduce memory use; supported native quantized compute is optional, while bf16 remains the default. Different checkpoint formats and attention backends may affect speed, memory use, or the exact image produced by a seed. See Models and quantization.
Control, paint, and identity editing may require their corresponding optional weights or adapters; selecting Raw or Turbo alone does not supply those editing capabilities.
What's New with Krea2_ComfyUI_Advanced
The supplied overview identifies v1.7.0 (November 2026) as its latest release. Its highlighted changes are:
- Hybrid Sigma Handoff: a new node hands a hybrid sampler from one method to another by sigma level rather than step count, and adds an SDE band per stage and a churn-form choice. This offers finer control over where the sampling behavior changes.
- Sweep Checkpoints: compare several Turbo checkpoints in one sweep folder and report, rather than assembling separate comparisons.
- More sweep choices: a trajectory probe and additional LoRA-picking options—including creator-based picks, one pick per prompt, and
#randomword options—make checkpoint and style comparisons more flexible. - Negative Guidance: a
negpip_negativefield provides another way to specify concepts you want to discourage. - Expanded guidance: documentation now covers hand-written sweep cells and the causes and controls of Turbo’s “wet effect.”
The documentation was also reorganized into focused guides. For updates beyond these highlights, consult the full change history.
Troubleshooting Krea2_ComfyUI_Advanced
My portrait looks oily, wet, or excessively textured
Turbo can over-emphasize skin highlights and fine texture. Try these changes one at a time so you can see which helps:
- For larger images, use
distilled_shift = resolutionrather than a low fixed shift. - Reduce steps if the image already looks complete; more steps are not always better.
- If you use a second-stage refine, try a gentler sigma curve or adjust the hybrid handoff.
- In Negative Guidance, try NAG text such as
wet skin, glossy skin, oily sheen, sweat, water droplets. NAG adds processing time. - Alternatively, try
negpip_negativefor similar concepts at lower cost. Avoid putting the same concept in both NAG and NegPiP, which can suppress it too strongly. - Describe the finish you want in plain language: “matte skin, natural texture, soft diffuse lighting.”
Do not use (matte skin:1.3) expecting standard prompt weighting: these nodes do not provide that kind of positive prompt weighting. See the negative-guidance guide.
My reference picture did not preserve the exact subject or layout
Check which reference method you connected. Vision Prompt provides semantic guidance—a picture can inform the description, but it is not a pixel-preserving edit. For a starting composition, encode the source with VAE Encode and connect its latent as init_latent. For the reference-latent edit pathway, use a compatible edit-trained LoRA.
A Reference Latents edit becomes a blocky mosaic
- Check whether an edit-trained LoRA is loaded.
- If not, disconnect
ref_latentsand use Vision Prompt or ordinary image-to-image instead. - If it is loaded, confirm that the LoRA is intended for this edit pathway. The base model alone cannot use reference latents safely.
An editing or guidance combination fails—or my negative guidance seems ignored
Inspect the connected Ultra sockets. Some methods cannot be combined because they need the same part of the generation process. Remove one conflicting method and rerun. If the console says guidance was skipped with style, paint, or reference latents, that is a documented limitation of those combinations, not evidence that your negative text was lost.
My prompt seems to ignore its ending
Krea 2 has a 507-token usable prompt budget. Ultra prints a token count and warns when words are dropped. Shorten long descriptions or LLM-expanded prompts, keeping the most important subject and visual instructions. As a practical guide, keep expanded English prompts under roughly 380 words. Vision Prompt checks its own budget and reports an error instead of silently cutting it.
The bottom of an image looks wrong, or an older seed no longer matches
The author reports a fix for an attention-mask issue that previously affected the bottom rows of Krea 2 images. With the corrected behavior, crop_bottom is no longer needed by default, and the same seed can produce a slightly different—but correctly rendered—image. If an older saved workflow still crops the bottom, check its stored crop settings rather than assuming the new defaults replaced them.
A checkpoint switch runs out of video memory
Try Eric Krea2 Unload Models to free VRAM before another heavy run, and lower the resolution if needed. The author reports fixes for earlier model-switch caching problems; if you are comparing behavior with an older pack version, consult the change history. Complex edit and reference combinations can also use substantial memory, so start around 1 MP when testing them.
Frequently asked questions
Do I need RES4LYF or another sampler pack?
No. This extension includes its own Krea 2 flow-matching samplers.
Should I use Turbo or Raw for a large image?
Choose Turbo with the 2× upscale decoder for a faster route. Choose Raw with multi-stage re-denoising when you want more control over refinement.
Is Vision Prompt the same as inpainting?
No. Vision Prompt influences the image’s described content or style. Use Paint with a mask when you want to regenerate a particular area.
Can I recover the settings used for a saved image?
Yes, if its generation settings were embedded in the saved PNG. Use Settings from Image to read them.
Will the same seed always give an identical image when I change performance settings?
Not necessarily. The author notes that an alternative attention backend can produce a related but non-identical image from the same seed. Model, sampling, and corrected attention behavior can also change results.
Learn More about Krea2_ComfyUI_Advanced
- Example workflow and group guide: the best starting point for exploring generation, reference images, and editing on one canvas.
- Image-to-image guide: Vision Prompt, source-image generation, depth and pose control, style, paint, and edits.
- High-resolution guide: choosing between progressive refinement and upscale decoding.
- Samplers and schedules: what the sampling choices do and how to pair them.
- Sweep guide: comparing settings visually and reading contact-sheet results.
- Negative-guidance guide: using NAG and NegPiP with Turbo.
- LoRA guide: searching, stacking, and using trigger words.
- Models and quantization: checkpoint formats, memory-saving compute options, and DeGrid.
- Full change history: release details beyond the highlights above.
