Name: Gladas' Workflow for Krea 2 Turbo
Rating: 5 (87 reviews)
Author: Gladas

OpenMOSS's new open-weights image model is in the generator. Text that reads right, layouts you control, prompts up to 4K tokens. Train your own LoRAs too.


MOSS Image: Text That Reads Right

Black Forest Labs' new open-weights image model is in the generator. Text-to-image, plus edits with up to 10 reference images. FLUX 3 Image is here too.


Flux.3 Klein: Edit With 10 References

Krea 3 is live in the generator. Add up to 6 reference images, describe the change, and pick Instant or Medium. Up to 2K output, from 30 Buzz per image, 33% off during launch week (through Oct 22).

Krea 3: Edit Any Image With a Sentence

HeyGen Video is live in the generator. Text, image or reference to video, 5 to 15 seconds, with dialogue, ambience and effects in every clip. 50% off through October.

HeyGen Video: Clips That Come With Sound

Nine Creator Score tiers, each with its own badge, plus achievements for publishing, reach, sales, judging and wins. Whatever you've already earned is waiting on [your journey page](/creators/journey).

Creator Journey: See How Far You've Come

The content half of the H3 Contest is open! Make a short video with H3, enter the crucible, the community judges it head-to-head. Over 900k in Buzz up for grabs.

H3 Content Contest: Six Seconds to Win

A training and content contest with over 2 Million Buzz in prizes and 50% off H3 generation for the duration of the contest.

The Civitai H3 Contest: 2 Million Buzz Up For Grabs

AIO Loaders have been removed as of v12. If you use AIO models, you can easily add a checkpoint loader to replace the existing loaders and connect them accordingly.The default Load Diffusion Model loader looks for your Krea 2 Turbo model in the models/diffusion_models folder.This workflow has the following functions/features:<ul><li>Text to Image</li><li>Multiple Lora Loading</li><li>Wildcards</li><li>Upscale</li><li>Detailers (removed on v12)</li><li>Save Image with full resource metadata in CivitAI readable format</li><li>This workflow is not setup to work with models that use CFG higher than 1.</li><li>These will work with Anima Turbo models, but you'll have to figure out the settings.</li></ul>Support Policy<ul><li>These workflows are provided as-is and without guaranteed support.</li><li>I do not provide support for third-party applications, modified versions, redistributed copies, reposts, bundles, or installations obtained from unofficial sources.</li><li>Inclusion of these workflows in another project does not imply my endorsement, involvement, or responsibility for that project.</li></ul>Disclaimer<ul><li>Responsibility for generated content falls on the user of these workflows.</li><li>Installation of ComfyUI and/or any related custom nodes for these workflows are at the installer's own risk.</li></ul>

This is a SamplerCustomAdvanced version of v12 with Prompt Reformat and SeedVR2.Important differences between v12 and v12b:<ul><li>You have manual control over the denoise setting on the 2nd pass. v12's KSampler (Advanced) nodes don't have that option.</li><li>The step count is combined between the blue Basic Scheduler node and the red Steps node.</li><li>If you are going to skip using the 2nd SamplerCustomAdvanced node and want your step counts to be accurate, make sure to set the red Steps value to 0.</li><li>Comfy Math had to be re-added to provide an accurate step count to Image Saver.</li></ul>This version uses the Basic Guider node instead of the CFG Guider node.<ul><li>If for some reason you need to use a negative prompt, you will have to replace the Basic Guider with a CFG Guider node and add a CLIP Text Encoder node for your negative prompt.</li></ul>

v12b

v12 is a VERY stripped down workflow. I started fresh with this workflow and the result is a faster (for me) and leaner workflow.<ul><li>Wildcard usage is still available along with BKWILDCARDS.</li><li>The "empty" latent source is the SDXL Empty Latent Image node from rgthree for simplicity.</li><li>This is setup as a two KSampler (Advanced) workflow with upscaling performed via latent upscale in the 2nd KSampler (Advanced) node.</li><li>Separate lora loaders are available for each KSampler and are set to match the color of the KSampler they are connected to.</li><li>Image Saver to save the image with CivitAI readable metadata. ComfyUI Manager doesn't seem to list the latest version of Image Saver. I recommend getting it from the project page.</li><li>A note with links to all the custom nodes used in this workflow can be found on the far left side.</li></ul>TL;DR of how the workflow works (these are the default settings):Pick the initial resolution on the SDXL Empty Latent Image node.<ul><li>This can be swapped for your preferred Empty Latent node. Just connect your latent output to the "latent_image" input on the blue KSampler.</li></ul>Set the Upscale Latent by "scale_by" setting to 2.00 to perform a 2x upscale. This can be adjusted as needed.I use the two KSampler (Advanced) to generate the initial image and proceed straight into the upscale.On both KSamplers:<ul><li>Steps: 10</li><li>CFG: 1.0</li><li>Sampler: euler</li><li>Scheduler: simple</li></ul>On the blue KSampler:<ul><li>start_at_step: 0</li><li>end_at_step: 6</li></ul>On the red KSampler:<ul><li>start_at_step: 6</li></ul>With these settings, I get a 2x upscaled image in around 35 to 40 seconds on a 5060ti 16GB with comfy kitchen enabled. Adding additional loras will likely increase the generation time.NOTE: You can choose to use only one KSampler. Just make sure to bypass the 2nd KSampler and the Upscale Latent by node.Removed:<ul><li>AIO Loader</li><li>LC123 nodes</li><li>Krea2 NegPip</li><li>Prompt Reformat</li><li>Detailer</li><li>SeedVR2</li></ul>

This is LC Speed v4 with some extras added:Loras used anywhere in the workflow should now be populated in the resource list when posting to CivitAI.Optional Latent Upscale node between the 2 SamplerCustomAdvanced nodes.<ul><li>This is intended for people not using the LC Speed Boost node, but want to upscale in latent space.</li></ul>Optional image upscale via the Upscale Image and Upscale Image (using model) nodes.

LC Speed v5

This is a SamplerCustomAdvanced version of v3, but with an optional 2nd pass.This is my version of the LC Speed test workflow's High Pass / Low Pass option. The main differences are:<ul><li>No special custom node for controlling the related settings.</li><li>Manual control over the denoise for the 2nd pass.</li><li>No detail daemon. IMO detail daemon is not needed for Krea2, but feel free to add it.</li><li>Separate lora loader for the 2nd SamplerCustomAdvanced.</li></ul>Added a few markdown notes explaining how I generally use the workflow.Version 3 is probably faster for most, but I have generally been partial to using SamplerCustomAdvanced in my workflows.

LC Speed v4

This version removes the SamplerCustomAdvanced nodes and returns to the LC Speed Boost KSampler.<ul><li>This results in a shorter, cleaner, faster, and easier to use workflow than v2.</li></ul>Added a text box for the Negative Prompt in case anyone uses this workflow for non-Turbo models.<ul><li>This workflow does work with Anima, but compatibility is not guaranteed with all Anima models.</li></ul>Added a quality of life (QoL) related node from the LC123 nodes for prompt cleanup when using an LLM or the included Prompt Reformat option. There are notes from the LC123 creator on how it works next to the node.

LC Speed v3

Experimental Workflow v2Added the option to use the High Pass/Low Pass version of the LC Speed setup.<ul><li>Added ModelSamplingAuraFlow based on how the LC Speed test workflow was setup.</li></ul>I had to get creative with the single KSampler version and convert it to a single SamplerCustomAdvanced version since Image Saver is a pain sometimes.<ul><li>The result is shared nodes for all the KSampler settings that will properly communicate with the Image Saver node.</li></ul>

LC Speed v2

This is an experimental workflow using a custom node created by <a target="_blank" rel="ugc" href="https://civitai.red/user/lonecatone23">lonecatone23</a> to speed up generation time. The article regarding this custom node can be found <a rel="ugc" href="https://civitai.red/articles/36132/speed-boost">here</a> on CivitAI. The custom node can be found on github <a rel="ugc" href="https://github.com/lonecatone23/ComfyUI_LC123_nodes">here</a>.This node allowed me to remove the upscaling groups from my workflow since I can generate a 2x SDXL resolution image in about 60% of the time.Removed the following:<ul><li>Latent Upscale</li><li>Ultimate SD Upscale</li></ul>Detailers are really not needed, but I left one in this version of the workflow just in case.

LC_Speed_v1

Krea 2 Turbo Workflow — v11Workflow changes<ul><li>Added an optional 2nd Pass Latent Upscale path with its own sampling settings.</li><li>Reworked resolution controls around a Width × Height preset and separate Scale by controls for initial generation and latent upscaling.</li><li>Added an upscale output switch. When both latent upscaling and USDU are enabled, the latent upscale result takes priority.</li><li>Removed the pre-USDU SeedVR2 stage, keeping SeedVR2 as an optional final cleanup pass.</li><li>Updated upscale and detailer model connections to use the NegPiP-patched model.</li><li>Updated workflow navigation and notes.</li></ul>Updated defaults<ul><li>Initial generation: Euler + Simple, 8 steps, CFG 1.</li><li>USDU: Euler + Simple, 3 steps, denoise 0.24.</li><li>USDU now uses 1536×1536 tiles with 128 padding. Full-image single-tile processing remains available by connecting the image dimensions to the tile-size inputs.</li><li>Face detailer denoise: 0.10.</li><li>Eye detailer denoise: 0.24.</li><li>Final SeedVR2 cleanup denoise: 0.10, with tiled VAE overlap increased to 128.</li><li>Both detailers and final SeedVR2 cleanup are bypassed by default.</li></ul>Notes<ul><li>The latent upscale target is calculated from the resolution preset, independently of the initial-image scale.</li><li>Upscale, detailer, and cleanup settings remain model-dependent—adjust them to suit your chosen finetune and hardware.</li></ul>

v10 — Workflow Cleanup and Separate LoRA Controls<ul><li>Simplified the processing chain by removing the second USDU pass and intermediate VAE encode/decode round trips.</li><li>Added separate LoRA loaders for initial generation, USDU, face detailing, and eye detailing.</li><li>Reordered the detailers to process the face first and eyes second, followed by final SeedVR2 cleanup.</li><li>Increased USDU tile padding to 256 and mask blur to 16. Connected image dimensions continue to provide a single full-image tile.</li><li>Updated both detailers to 3 steps, with guide size and maximum size set to 1536.</li><li>Adjusted eye detection sensitivity and mask filtering, and added an eye-specific detailer prompt.</li><li>Simplified group controls and retained bypass paths for optional detailers and final cleanup.</li><li>Cleaned up the layout and added a bookmark for Initial Image Settings.</li><li>Updated the prompt reformatter instructions and left the main prompt empty for users to enter their own.</li></ul>My default settings are not the bible of Krea 2 + ComfyUI. Please adjust them to fit your needs.

Krea 2 Turbo Workflow v9This update expands the refinement stages, configures USDU for full-image tiles, and improves routing between optional processing stages.<ul><li>Full-image USDU tiles: Both USDU passes automatically use the incoming image’s width and height for their tile dimensions. Disconnect these inputs to set smaller tiles manually.</li><li>Added a second USDU pass: Includes independent sampler, step, denoise, and LoRA controls, plus a before/after image comparer. The first pass uses 3 steps at 0.22 denoise; the second uses 2 steps at 0.15 denoise.</li><li>Added VAE encode/decode stages: Images pass through a VAE round-trip before USDU 2 and before each detailer. All of these stages follow the shared VAE Switch.</li><li>Updated detailer defaults: Both detailers now use 2 steps at 0.15 denoise and are enabled by default.</li><li>Updated image routing: Added fallback image switches after USDU 2 and each detailer. Detector inputs now match the images processed by their respective detailers.</li><li>Enabled post-USDU SeedVR2 by default.</li><li>Updated scheduler defaults: Initial generation and both USDU custom schedules use Normal. Detailers retain Simple.</li><li>Updated wildcard handling: Replaced the BK Wildcards selector and added conditional period handling before its text is concatenated with other prompt content. Empty text remains empty.</li><li>Adjusted defaults and cleanup: Selected VXP Krea2 beta3 INT8, cleared the NegPiP text, turned off the USDU/detailer LoRA switches, removed the unused Power Puter node, and revised workflow notes and labels.</li></ul>The second USDU pass is an optional refinement experiment and can improve or worsen an image. Compare the results and bypass it when it does not help. Full-image tiles can also increase memory use and processing time; settings remain model- and hardware-dependent.

Krea 2 Turbo Workflow — v8v8 streamlines the finishing stages, updates the optional Qwen prompt rewriter, and improves prompt concatenation.Workflow changes<ul><li>Removed one detailer, leaving the face and eye detailers as the two optional detailer stages. These, can be used for any detection model, but are currently set for face and eyes.</li><li>Removed Easy Color Correction and its comparison preview. ComfyUI-EasyColorCorrector is no longer used by this workflow and has been removed from the dependency note.</li><li>Moved the optional post-USDU SeedVR2 pass after both remaining detailers, so it processes their combined result when enabled.</li><li>Added a final image switch that feeds the final comparison and image saver from either the post-processing SeedVR2 result or the preceding image.</li><li>Added spaces between the combined positive prompt and wildcard text so descriptions do not run together.</li><li>Removed the outdated note instructing users to manually connect wildcard outputs; the wildcard branches are already wired through their switches.</li></ul>Sampling and detailer defaults<ul><li>Changed the initial-generation scheduler from Normal to Simple, retaining Euler, 8 steps, and CFG 1.</li><li>Changed USDU from Normal to Simple, retaining 3 steps, CFG 1, and 0.22 denoise. The saved setup uses 1024 × 1024 tiles with 128 padding. Full-image refinement remains an optional manual configuration described in the workflow note.</li><li>Enabled the USDU “Use loaded LORA models?” switch by default.</li><li>Updated the face detailer to Euler + Simple, 3 steps, and 0.40 denoise. Increased the face detector threshold from 0.75 to 0.80 and MaskToSEGS drop size from 10 to 15.</li><li>Updated the eye detailer to Euler + Simple, 3 steps, and 0.30 denoise. Cleared the previous <code>eye</code> wildcard text.</li></ul>Optional prompt rewriter<ul><li>Revised the system instructions to clarify input priorities, attribute ownership, clothing coverage and layering, prop handling, and the preservation of different rendering styles across the subject. Feel free to change them to your own instructions.</li><li>Kept camera details near the beginning of the requested output structure and tightened the instructions for producing only the finished prompt.</li><li>Increased the text generator’s repetition penalty from 1.00 to 1.05.</li><li>Replaced the previous Think Guard and Trim nodes with a regex cleanup node targeting a leading “Final Prompt” label.</li><li>Connected the prose preview directly to the generator output, allowing its raw text to be inspected before label cleanup.</li></ul>Saved configurationThe default image path is 832 × 1216 generation, 2× Lanczos resizing to 1664 × 2432, and tiled USDU refinement. Wildcards, Qwen rewriting, NegPiP, both detailers, both SeedVR2 branches, and the attention-backend override are bypassed in the supplied file. The positive prompt is left blank for users to enter their own prompt.Settings are starting points for the selected model and can be adjusted to suit it.

Krea 2 Turbo Workflow - v7This update focuses on prompt handling, wildcard integration, and revised upscale defaults.Prompt reformatting<ul><li>Reworked the optional Qwen Prose Reformat setup to use a dedicated, editable system-prompt node, replacing manually assembled chat wrappers.</li><li>Expanded the instructions for clothing coverage, layering, footwear, prop handling, and physical relationships.</li><li>Added guidance for preserving mixed styles, such as an anime face with realistically rendered hair, body, and clothing.</li><li>Added instructions for supplying a suitable background and cinematic lighting when those details are unspecified.</li><li>Increased the text-generation limit from 450 to 1,024 tokens and adjusted generation settings.</li><li>Added a troubleshooting note for failed reformat attempts.</li></ul>Wildcard integration<ul><li>Connected Impact Wildcards and BKWILDCARDS through dedicated switches with empty-text fallbacks.</li><li>Both wildcard sources now combine with the main positive prompt before optional reformatting. The combined prompt also feeds generation when reformatting is bypassed.</li><li>Added separate text previews for both wildcard sources.</li><li>Updated the BKWILDCARDS Selector from 0.9.18 to 0.9.19.</li><li>Both wildcard sections remain bypassed by default.</li></ul>Upscaling and refinement<ul><li>Revised USDU defaults to Euler + Normal, 3 steps, CFG 1.0, and 0.22 denoise.</li><li>Changed USDU tiles from 1536 × 1536 to 1024 × 1024, retaining 128-pixel padding. Tiles are more friendly for folks with non-whale hardware, but a large single tile is generally more stable.</li><li>Set USDU to run without the loaded LoRAs by default. The existing switch allows them to be enabled.</li><li>Adjusted optional SeedVR2 denoise settings to 1.0 before USDU and 0.15 after USDU.</li><li>Changed SeedVR2 tiled VAE encoding and decoding to 1,024-pixel tiles with 64-pixel overlap.</li><li>Adjusted the third detailer’s denoise to 0.25 and set its wildcard specification to <code>eye</code>. You can change to something else or delete the text. When I use this detailer, it is only for eyes.</li></ul>Other changes<ul><li>Changed the default portrait resolution to 832 × 1216.</li><li>Color correction is now bypassed by default.</li><li>Removed the Global Seed node and its group control.</li><li>Updated seed-locking instructions, upscale recommendations, and workflow labels.</li></ul>Setup notesUpdate BKWILDCARDS for the revised selector. Wildcards, prompt reformatting, SeedVR2 passes, detailers, and color correction are bypassed in the supplied workflow and can be enabled as needed.

Krea 2 Turbo Workflow v6This update fixes the NegPiP bypass issue, adds an optional cleanup pass after USDU, and improves prompt visibility.Fixes and improvements<ul><li>Fixed NegPiP prompt text affecting generation while bypassed. Replaced the previous prompt input with a multiline text node and a switch that supplies an empty value when the NegPiP group is bypassed.</li><li>Added an actual prompt preview. Displays the combined text being sent to CLIP Text Encode, making it easier to check whether NegPiP text is included.</li><li>Added an optional post-USDU SeedVR2 pass. Intended to help clean up residual noise after USDU, with its own before-and-after comparison. Disabled by default.</li><li>Reworked the existing pre-USDU SeedVR2 pass to use SamplerCustomAdvanced with separate sampling controls.</li><li>Added guidance for using USDU as a single-tile second pass, plus an image-size node before USDU to make connecting the tile dimensions easier. This setup is optional and is not connected by default.</li></ul>

Krea 2 Turbo Workflow — v5 Release Notes<ul><li>Added optional NegPiP support, with a dedicated prompt field and an instructions note. Enter negatively weighted terms using the format <code>(unwanted elements:-1)</code>. The NegPiP nodes are bypassed by default. If you're seeing some weird behavior while it is bypassed, remove the text contained in the NegPiP prompt node.</li><li>Integrated NegPiP into the model and prompt connections, including the existing LoRA selection paths for upscaling and detailing.</li><li>NegPiP's prompt box text is not saved into the positive prompt. The reason for this is to prevent false flag TOS violations when the intent is to use it as a negative prompt.</li><li>NegPiP is not connected to the Prompt Reformat in order to not confuse it. The text will be added from the NegPiP prompt before it is sent to the CLIP Text Encode node.</li><li>NegPiP for Krea 2 can be found here (<a target="_blank" rel="ugc" href="https://github.com/blue-pen5805/ComfyUI-krea2-negpip">https://github.com/blue-pen5805/ComfyUI-krea2-negpip </a>) and in ComfyUI Manager.</li><li>Updated the custom-node dependency list to include ComfyUI-krea2-negpip.</li><li>Made Qwen prompt reformatting optional by default. The reformatting section is now bypassed on load, allowing the standard prompt to pass through directly.</li><li>Adjusted detailer settings: face and eye detailer denoise increased from 0.22 to 0.24, and the hair detailer’s mask conversion now uses combined regions. Detailers remain bypassed by default.</li><li>Disabled Global Seed by default, leaving individual seed controls available.</li><li>Renamed “Basic Upscale” to “Pre-USDU Upscale” to clarify its role.</li><li>Refreshed the included example prompt.</li><li>Added a few more markdown notes. These briefly cover how to connect the wildcard nodes and how NegPiP works.</li><li>Added individual steps and denoise nodes to the USDU group to sync the values on BasicScheduler and the USDU node. This is purely cosmetic, but should alleviate any confusion about the values being different.</li><li>Did some minor cleanup of the layout and also pulled out (not deleted) the Image Comparer nodes out of their respective groups. For some reason, they were reactivating the groups while bypassed.</li></ul>

Krea 2 Turbo Workflow — v4 Release NotesThis update reorganizes the workflow, adds optional prompt rewriting and attention controls, and revises the default generation and upscale settings.New additions<ul><li>Added an optional Qwen Prompt Reformat section to rewrite prompts into natural-language descriptions. Includes instructions to preserve the supplied details, text cleanup, and an output preview. Bypassed by default. Credit goes to <a target="_blank" rel="ugc" href="https://civitai.red/user/bkidderz">@bkidderz</a> for this. Note: the output sometimes breaks on this, but overall it is pretty reliable. This is very similar to using Prompt Enhancer. It will add about 20 seconds to your generation time depending on your hardware.</li><li>Added the <a target="_blank" rel="ugc" href="https://civitai.red/models/2849795/bkwildcards-custom-node-for-comfyui">BK Wildcards Selector</a> as an additional prompt-building tool. The selector is currently bypassed and unconnected; its output can be connected into one of the positive prompt text boxes.</li><li>Added an optional Model Attention Backend node with Comfy Kitchen Attention selected. Comfy Kitchen is natively part of ComfyUI and does not require you to install anything new unless your Comfy install is very old. Bypassed by default.</li><li>Added accompanying notes and updated the custom-node list.</li></ul>Workflow changes<ul><li>Reorganized groups and bypass controls for model loading, wildcards, upscaling, detailers, and image saving.</li><li>Removed the separate Upscale with Model branch. Basic Lanczos resizing, SeedVR2, and Ultimate SD Upscale remain available.</li><li>Enabled SeedVR2 in the saved workflow and changed its denoise setting from 0.40 to 0.70.</li><li>Changed LoRA routing so the Power LoRA Loader patches the diffusion model while text encoding uses the loaded text encoder directly.</li><li>Replaced zeroed positive conditioning with a separate, empty negative-prompt encoder.</li><li>Custom latent dimensions can now be entered directly without the resolution calculator overriding them by default.</li><li>Updated several explanatory notes throughout the workflow.</li></ul>Updated saved defaults<ul><li>Initial resolution: 832 × 1216 → 768 × 1344.</li><li>USDU tile size: 1280 × 1280 → 1536 × 1536.</li><li>USDU tile padding: 96 → 128.</li><li>Global Seed behavior: increment for each node → randomize for each node.</li><li>One optional detailer’s selected detector changed from hands to hair.</li></ul>The workflow retains 2× basic upscaling, optional detailer passes, color correction, image comparisons, and image saving with metadata. Select your installed models and review the enabled groups before generating.

Added SeedVR2 upscaling to the workflow.No new custom nodes are needed, but you will need the related model and VAE to use them.<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/SeedVR2/resolve/main/vae/seedvr2_ema_vae_fp16.safetensors">seedvr2_ema_vae_fp16.safetensors</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/SeedVR2/resolve/main/diffusion_models/seedvr2_3b_int8_convrot.safetensors">seedvr2_3b_int8_convrot.safetensors</a></li></ul>

Added AIO model compatibility<ul><li>AIO models have the Checkpoint, VAE, and CLIP in a single file.</li><li>AIO models need to be placed in your models/checkpoints folder to use this loader.</li><li>This loader is in its own group. If both groups are enabled, it will default to the 'normal' loader instead of this one.</li></ul>USDU<ul><li>I disconnected the tile_width and tile_height inputs and have set them to 1536 x1536.</li><li>Since there is no real guidance on how Krea 2 Turbo behaves with USDU, the settings adjustments were made with the help of ChatGPT after some discussion today.</li><li>Adjust the settings to your preference.</li><li>Added a new markdown with some ChatGPT recommendations for USDU settings.</li></ul>Workflow default settings<ul><li>The first 20 sample images were generated using the default settings on the saved workflow.</li><li>Depending on what model and loras you use, these settings will need to be adjusted. Some models just perform bad with USDU no matter what. In those cases, it's better to generate a higher res initial image and either bypass USDU or set the upscale to 1.0 and use USDU as a 2nd pass while setting the tile width and height to match the resolution of the initial image.</li></ul>

Initial release.Some minor changes from my old workflow (if you were using it).Removed CFG Zero Star<ul><li>This does nothing at CFG 1.</li></ul>Removed the text box for the Negative Prompt<ul><li>Negative prompts are useless at CFG 1, so this is just decluttering.</li></ul>Image Comparers now show the 'after' result by default.

Gladas' Workflow for Krea 2 Turbo

2026-10-01-161351_vxpKrea2Nsfw_beta3INT8_998123531747912.png

wide hips

downblouse

pg-13

suggestive

oral invitation

pg13

sexy

huge breasts

thick thighs

physical violence

weapon violence

revealing clothes

convenient censoring

corpses

male nudity

female swimwear or underwear

female nudity

breast out

male underwear

hair over breasts

female swimwear

gigantic breasts

graphic violence or gore

emaciated bodies

nude