CivArchive

    🟢 Muse by Stable Yogi Ā· Krea 2 V3.0 — Extended and Extended Edit

    Two upgraded cuts of Muse V3.0, and both are free for the community.

    Same Muse look you know, now with more room to work: run more steps, push the guidance a little, and edit your images right inside the model.

    ## What's in the pack

    šŸ‘‰ Get the free Extended versions → HERE

    No cost, just the code. Drop a ā¤ļø and tag me if you make something you love.

    🟣 Extended (Turbo)

    The full Muse V3.0 look with the brakes off. The standard turbo model is tuned for quick 8 step images at CFG 1. Extended opens that up so you can slow down for a cleaner, richer result and give the prompt more pull.

    🟣 Extended Edit

    Everything in Extended, plus image editing built in. Give it a picture and a plain instruction: swap an outfit, change a face or hair, add or remove someone, extend the frame, fill a gap, or try on a different look. It keeps the person looking like themselves.

    ## Settings (both cuts)

    - Steps 8 to 20 (16 is the sweet spot) Ā· CFG 1 to 3 (try 2) Ā· Euler Ā· Simple.

    - Keep CFG at 3 or below. Higher pushes it too hard and the image starts to break.

    - Extended Edit runs with the comfyui-krea2edit node pack. The editing is baked in, so skip the separate edit LoRA loader.

    - Model only. Pair with the Qwen3-VL text encoder + Qwen image VAE, same as base Krea 2.

    ## šŸŽ How to get it (free)

    Both cuts are free with coupon *EXTENDED-MV3** at checkout.

    Editing is powered by the Krea 2 Identity Edit LoRA by conradlocke (Krea 2 Community License). Thanks to them.

    ===========================================================

    🟢 Muse by Stable Yogi · Krea2 · V3.0

    The next step for Muse — the editorial look you know, now with more striking presence.

    ✨ What's NEW in V3.0

    - Everything from V2.5, elevated. The full Muse blend plus a more striking, magazine-grade edge.

    - Standout faces. Editorial beauty, believable and polished — holds shot to shot.

    - Glossy, cinematic light. Dimensional, produced, that Muse signature.

    - Handles anyone. Young to mature, exactly as prompted.

    ⭐ Want more? The PRO version pushes presence and depth further — see below.

    Made by Stable Yogi 🧘 — drop a ā¤ļø if Muse is your look.

    Muse by Stable Yogi Ā· Krea2 Ā· V2.5 NVFP4

    MuseĀ byĀ StableĀ YogiĀ Ā·Ā Krea2Ā Ā·Ā V2.5 — NVFP4

    TheĀ NVFP4Ā (4-bit) build of Realism V1.0 — made forĀ RTX 50-series / BlackwellĀ cards.

    • Smallest & fastest — ~8 GB, native FP4 tensor cores. Lowest VRAM, top speed on 50-series.

    • Same realism — real skin, natural light, true-to-life — no quality loss from 4-bit.

    • Runs in ComfyUI + the latest ForgeĀ (Blackwell only — on older cards use the INT8 / fp8 / GGUF versions instead).

    āš™ļø Settings:Ā 8 steps Ā· CFG 1 Ā· Euler Ā· Simple Ā· native 1024²+
    (Model only — pair with the Qwen3-VL text encoder + Qwen-Image VAE, same as base Krea 2.)

    Made by Stable Yogi , get Pro release on my portal.

    Muse by Stable Yogi Ā· Krea2 Ā· V2.5 EXTENDED

    🟢 MUSE V2.5 EXTENDED

    YourĀ settings,Ā yourĀ look — MuseĀ V2.5Ā withĀ roomĀ toĀ move.

    Most turbo models pin you to one setting: change the steps or the guidance and the image falls apart.Ā Extended doesn't.Ā ItĀ holdsĀ togetherĀ acrossĀ aĀ realĀ workingĀ range,Ā soĀ youĀ canĀ tuneĀ theĀ resultĀ insteadĀ ofĀ acceptingĀ it.

    āš™ļø Settings

    • Recommended: 16 steps Ā· CFG 2 Ā· Euler Ā· Simple Ā· native 1024²+

    • Works across: 8–20 steps Ā· CFG 1–3

    • āš ļøĀ Don't go above CFG 3.

    ✨ Why you'd want it

    • Dial in your own look.Ā Lower CFG for soft and natural, higher for punch and contrast. Fewer steps for speed, more for polish. Your call, not the model's.

    • Extra steps actually pay off.Ā At 16 steps you get the richest version of Muse — deeper light, warmer tone, more presence.

    • It follows your prompt harder.Ā More guidance headroom means when you ask for a specific scene, you get that scene.

    • Still quick when you're in a hurry.Ā 8Ā stepsĀ worksĀ fineĀ forĀ draftsĀ andĀ batchĀ runs.

    šŸŽÆĀ WhichĀ oneĀ shouldĀ IĀ use?

    BestĀ for Muse V2.5Ā (standard)Speed.Ā SharpestĀ atĀ 8Ā steps — theĀ everydayĀ default.

    Muse V2.5 Extended

    ControlĀ andĀ qualityĀ ceiling.Ā GiveĀ itĀ 16Ā steps,Ā CFGĀ 2.

    šŸ“¦ Download — INT8-ConvRotĀ (fast INT8 tensor cores, loads natively in Forge + ComfyUI,

    MadeĀ byĀ StableĀ YogiĀ šŸ§˜Ā ā€”Ā dropĀ aĀ ā¤ļøĀ andĀ showĀ meĀ whatĀ youĀ tuneĀ itĀ to. Looking for a Pro Version check here on my portal.

    =========================================================

    Muse by Stable Yogi Ā· Krea 2 Ā· V2.5

    The full-spectrum edition of Muse on Krea 2 — same photoreal turbo, a wider, richer blend.

    ✨ What's NEW in V2.5

    - A broader, more varied blend — more range across faces, skin, and lighting moods.

    - True photorealism — real skin texture, pores, freckles, natural light (not AI-glossy).

    - Believable faces, consistent — holds up shot to shot.

    - Turbo — full images in 8 steps.

    šŸ”§ V2.0 vs V2.5

    - Companion release to V2.0 — we put out both so you can pick your favourite. V2.0 is the tighter, more focused cut; V2.5 is the fuller blend with a touch more variety. Same engine, same speed.

    šŸ“¦ Downloads

    - INT8-ConvRot — fast (INT8 tensor cores), loads natively in Forge + ComfyUI.

    - Q8 GGUF — near-lossless. Ā· Q4 GGUF — low VRAM / small download.

    - (Model only — load the *Qwen3-VL text encoder + Qwen-Image VAE** alongside, same as base Krea 2; they auto-download.)*

    āš™ļø Settings

    - 8 steps Ā· CFG 1 Ā· Euler Ā· Simple Ā· native 1024–2048.

    ⭐ Want more punch? Grab the PRO edition from my portal, dials detail and depth up further.

    Made by Stable Yogi — drop a ā¤ļø and tell us: V2.0 or V2.5?

    =========================================================

    Muse by Stable Yogi — Krea2 Ā· V2.0 EDIT (INT8)

    For details of this release, please read the "About this model" section

    ==========================================================

    Muse by Stable Yogi Ā· Krea2 Ā· V2.0 (INT8)

    The next evolution of Muse on Krea 2 — more real, more detailed, same turbo speed.

    ✨ What's NEW in V2.0

    • A real jump in photorealism — skin now has genuine texture: pores, freckles, fine detail (less "AI-smooth").

    • Shot-on-camera look — believable faces, natural lived-in complexion.

    • Richer, more dimensional lighting — window light, soft shadows, real depth.

    • Holds up across prompts — lifelike skin and faces, consistent shot to shot.

    šŸ”§ What CHANGED (vs V1.5)

    • Built on the V1.5 engine, evolved with fresh photoreal character detail.

    • Cleaner, more natural rendering — over-polished feel gone.

    • StillĀ turbo — full images inĀ 8 steps.

    šŸ“¦ Downloads — INT8-ConvRotĀ (Forge + ComfyUI native) Ā·Ā Q4 GGUFĀ (low VRAM).Ā (Load Qwen3-VL TE + Qwen-Image VAE alongside — auto-download.)
    āš™ļø Settings — 8 steps Ā· CFG 1 Ā· Euler Ā· Simple Ā· native 1024–2048.
    ⭐ PRO edition on my portal dials it up further.

    ===========================================================

    Muse by Stable Yogi — Krea 2 Ā· V1.5

    Photoreal Krea 2 — fast, versatile, built to look genuinely real.

    ✨ What it is

    • Photorealistic model onĀ Krea 2, tuned by Stable Yogi Ā·Ā turbo (8 steps)Ā Ā· plug-and-play with the Krea 2 TE + VAE.

    šŸŽÆ Great for

    • PhotorealĀ portraitsĀ &Ā full-body

    • Any character

    • Fashion & editorialĀ 

    • Lifestyle & scenesĀ 

    • better lighting

    • Strong prompt-followingĀ 

    • Natural skin & facesĀ 

    šŸ“¦ Downloads — Q8 GGUF (recommended)Ā Ā·Ā Q4 GGUF (low VRAM)Ā Ā·Ā fp8 (compat)Ā (load TE + VAE alongside)
    āš™ļøĀ 8 steps Ā· CFG 1 Ā· Euler Ā· Simple Ā· native res
    ⭐ Want more punch? The PRO edition on my portal is dialed up further.

    ===========================================================

    Muse by Stable Yogi — Krea 2 Ā· V1 Experimental

    A photorealistic model onĀ Krea 2, tuned for lifelike portraits, natural skin & light — at turbo speed.Ā V1 Experimental: an early release, feedback welcome.


    ✨ Photoreal portraits & characters · 8-step turbo · native high-res · versatile (portrait / editorial / lifestyle / concept) · plug-and-play (TE + VAE built in).


    āš™ļøĀ 8 steps Ā· CFG 1 Ā· Euler Ā· Simple Ā· native 1024–2048.


    šŸ“¦ Fp8 (full quality) Ā· Q8_0 / Q4_0 GGUF (low-VRAM).

    Looking for a Pro version get it free at my portal


    Drop a ā¤ļø and share your creations.

    To run this model on Forge UI, look here https://civarchive.com/articles/32010/run-krea-2-in-forge-free-extension-first-on-forge

    Description

    Experimental Build with some spice.

    Looking for a Pro version get it free at stable yogi

    FAQ

    Comments (29)

    DrakeniJun 30, 2026Ā· 2 reactions
    CivitAI

    LOL how did you manage to INCREASE the size of a model by training it. This is like 50% larger than Krea 2 fp8...

    willywizardJun 30, 2026Ā· 4 reactions

    Because its not a fine-tune or what you would normally consider to be a "trained model", most of the "trained models" are barely trained, with a bunch of Loras baked in

    DrakeniJun 30, 2026

    @willywizardĀ That isn't how checkpoint training works. You could train it with a ton more data and not increase the size of the model. We aren't increasing the parameters with normal training.

    Stable_Yogi
    Author
    Jun 30, 2026Ā· 2 reactions

    The fp8 is a completeĀ all-in-one — it packs the full image model, the prompt text-encoder, and the VAE into one file withĀ zero wasted space, so you just load it and generate, nothing else to download. That's genuinely why it's ~20 GB. If you want a smaller download, theĀ Q8 GGUFĀ is leaner and actually a touch sharper — you load the text encoder + VAE alongside it.

    willywizardJun 30, 2026

    @Stable_YogiĀ Oh... Eww, Okay yeah I didn't see that it said "AIO" lmao

    ferrrett33Jun 30, 2026

    @Stable_YogiĀ Correct me if i'm wrong but for most of these more advanced models we have to load the text encoder and vae seperately anyway. So if we have a workflow already set up for that, why the all in one?

    Light7799Jun 30, 2026

    @ferrrett33Ā It's user friendly for anyone new to the scene of image gens, especially locally genned. They might just want to plug and play, and not have to add a node text encoder, vae, connect them etc. or might not even know how. I would say even a good bit of people don't know clip is a text encoder.

    ferrrett33Jun 30, 2026

    @Light7799Ā Hmm, for who exactly, even the ComfyUI workflow for Krea 2 includes the TE and VAE loaded separately. The problem with this is if we have multiple models they take up insane amounts of disk space and no way to trim them. AIO's haven't been a thing really, if you look at hundreds of models here on this website beyond SD and SDXL I mean.

    DrakeniJul 1, 2026

    @Stable_YogiĀ GGUF are slow.

    aising23Jul 1, 2026

    @ferrrett33Ā You can use the same workflow. Just disable and bypass the TE and VAE nodes. If you experiment with many models, then of course it's best to keep them apart to save space. But AIO is a better choice if you stick to few models. Pre-merging LoRa into a base model also works better, instead of doing calculations on the fly.

    bielzimff00187Jun 30, 2026Ā· 22 reactions
    CivitAI

    With this quality.. I assume the data was AI-generated... What a waste of time and the model's potential....

    novice286479Jun 30, 2026Ā· 1 reaction
    CivitAI

    Your website does not allow completing a download on slower connections. Even 9GB did not complete in 15 min. Wasted bandwidth.

    Stable_Yogi
    Author
    Jul 3, 2026

    There might be some network issues at that time , but everything works fine, You can retry on small models and confirm. If you face any issues ping me on chat.
    Thanks

    byvalyiJun 30, 2026Ā· 1 reaction
    CivitAI

    It doesn't even suck. Welcome to Pony

    ProtoniumJun 30, 2026Ā· 19 reactions
    CivitAI

    It is a weird thing to say on an AI website, but I think this is genuinely best described as AI slop.

    skyrimer3dJun 30, 2026Ā· 3 reactions
    CivitAI

    ok this is my new default krea 2 model, it works great, i only wish there was a INT8 version too.

    Stable_Yogi
    Author
    Jul 3, 2026Ā· 1 reaction

    Try the Q8 on version 1.5.

    lhy666Jul 5, 2026Ā· 1 reaction

    @Stable_YogiĀ his talking about the new int8_convrot quantization. not gguf 8bit

    jimzlfJul 7, 2026Ā· 1 reaction

    @Stable_YogiĀ gguf is much slower

    Stable_Yogi
    Author
    Jul 7, 2026Ā· 4 reactions

    @jimzlfĀ Yeah that's normal with GGUFs, i will prepare the INT8 and upload soon.

    ogbtjyqat450Jul 1, 2026Ā· 2 reactions
    CivitAI

    Someone have a wf for this? Also: What does the pro version have over this one? Thanks.

    Stable_Yogi
    Author
    Jul 3, 2026Ā· 3 reactions

    Pro version is more vibrant with extra data and refining. Currently Krea2 Pro versions are free, you can grab here https://stableyogi.com

    delta45424155Jul 1, 2026Ā· 1 reaction
    CivitAI

    All other krea2 models and aio work but yours gives me this error:
    RuntimeError: mat1 and mat2 shapes cannot be multiplied (42x2560 and 4096x2560)

    if i load clip model separately it works.

    Stable_Yogi
    Author
    Jul 1, 2026Ā· 1 reaction

    I have removed the AIO and added the FP8 without the Clip , Try that.

    ogbtjyqat450Jul 2, 2026Ā· 1 reaction

    @delta45424155 which clip model and vae did you use?

    delta45424155Jul 2, 2026Ā· 1 reaction

    @ogbtjyqat450Ā standard ones for any Krea2

    Ses_AIJul 1, 2026Ā· 5 reactions
    CivitAI

    What is second file is for?

    Stable_Yogi
    Author
    Jul 3, 2026Ā· 1 reaction

    Which file , please be specific ?

    Ses_AIJul 3, 2026Ā· 3 reactions

    @Stable_YogiĀ at the moment of commenting there was a 400mb file, I see you uploaded checkpoint iinsted. So nvm I guess