🚀 Wellness Physique V6 (The 2MP LoKR Run)
Trained on RunPod via Fizgig, pushing 286 native 2MP high-resolution source images through a LoKR (Rank 4) architecture across 10 epochs (6 hours, 32 minutes total training time).
This version delivers an unprecedented level of anatomical density, vascularity, and skin-level micro-detail.
🔥 What’s New in V6:
Native 2MP Pipeline: Trained at true 2MP on crystal-clear source images. The jump in micro-detail—hamstring-glute tie-ins, quad separation, striations, surface vascularity, and pores—is a massive upgrade over standard 1MP runs.
Natural Language Captions: No rigid comma-separated tag soup. The LoKR responds dynamically to conversational sentences, giving you full creative control over outfits, environment, and muscle conditioning without fighting the weights.
LoKR Architecture: Delivers deep anatomical structure and believable density without the plastic, over-baked look common with high-rank standard LoRAs.
📦 The Checkpoints (Fitness vs. Wellness):
🏋️ 1. Fitness Edition (Epoch 4)
Aesthetic: Toned, athletic, commercial fitness physique.
How to Prompt: Minimalist. Just include the trigger word (wellness_physique). You don't need to add extra muscularity or vascularity cues—it is dialed in to deliver an athletic look out of the box.
🏆 2. Wellness / Bodybuilding (Epoch 6, Epoch 9 & Final Epoch 10)
Rather than a linear progression of "light to heavy," these checkpoints settled with distinct anatomical tendencies. Depending on your phrasing and angle, each reacts differently:
Epoch 6, 9, and Final (10) have their own unique strengths. For example, depending on your prompt, Epoch 6 can actually produce sharper definition and deeper separation on backshots and posterior chain framing than the later epochs, while Epoch 9 and 10 might emphasize frontal quad sweeps or overall mass differently.
They aren't "better or worse"—they are distinct variations. I strongly recommend testing the same seed across all three to see which checkpoint reacts best to your specific scene.
✍️ How to Prompt (Without Templates)
Because this model was trained on conversational language rather than rigid tags, you do not need a strict prompt template. Instead, write in fluid, descriptive phrases.
Think of your prompt as building blocks: describe what you want to see, and omit what you don't. If you want a softer look, leave out the vascularity and striation cues.
Here is the exact vocabulary the LoKR recognizes from the training data:
1. Conditioning & Muscular Detail
Vascularity: abdominal vascularity, vascularity across the quadriceps and forearms, fine surface veins
Striations & Separation: visible striations, fine muscular striations, deep quad cuts, glute-hamstring separation
Finish & Tan: deep competition tan, even stage tan, glossy skin sheen under warm directional lighting, matte skin finish
2. Posing & Angles
Rear & Posterior: viewed from the rear, rear three-quarter view, rear-facing pose showcasing gluteal contours and hamstring depth, leaning forward at the hips (Great for Epoch 6)
Frontal: front-facing stance with hands resting on hips, standing upright in a front-facing pose, arms raised overhead with hands clasped
Profile & Turn: side profile showcasing glute projection and quad depth, quarter-turn pose with one leg slightly bent
Candid / Casual: mirror selfie holding a smartphone, flexing arm in front of a mirror, seated on an exercise bench
3. Outfits & Scenery
Stage & Competition: sparkly iridescent competition bikini, beaded bikini with rhinestone straps, clear high-heeled platform sandals, dark stage backdrop with directional spotlights
Gym & Athletic: ribbed crop top, high-waisted seamless shorts, corset waist trainer, knee-high socks and athletic sneakers, gym interior with squat racks, dumbbells, and industrial overhead lighting
Casual & Lifestyle: frayed denim shorts pulled down at the hips, oversized faded t-shirt tied at the waist, sauna with vertical wood paneling, patio deck with lush tropical greenery
(Check the uploaded showcase images to see exact examples of how these phrases come together in full prompts!)
💪 Women's IFBB Wellness & Superheroine Muscle LoRA
This LoRA is tailored specifically for Women's IFBB Wellness & Female Physique aesthetics. It adds realistic anatomical female muscle definition, vascularity, and muscle striations focused on a bottom-heavy female muscularity approach.
Inspired by the Women's IFBB Wellness division, it emphasizes heavy lower-body female muscle development (massive glutes, quads, hamstrings, and calves) while maintaining a lighter, athletic upper body.
Perfect for Women's Wellness renders, female fitness/physique models, or female superheroine character designs.
🚧 Current Status & Future Updates
While this model is not 100% perfect yet, it handles photorealism and natural skin textures exceptionally well.
Known Areas for Fine-Tuning: Back muscle definition and abdominal details can sometimes be less consistent depending on the prompt and base checkpoint.
Future Roadmap: I am actively working on subsequent iterations to improve back and abdominal definition, refine dataset captions, and release updated versions.
Your feedback, test renders, and critique in the comments are hugely appreciated to help shape v4!
💎 100% Real Dataset & Training Philosophy
NO AI-Generated Images: Trained exclusively on 100% real-world photography of real female IFBB Wellness athletes. No AI-on-AI training degradation.
Athlete Privacy & Anonymization: To respect the privacy of the athletes, all dataset images were fully anonymized (faces cropped/removed or anonymized) during dataset preparation.
High-Care Dataset (70 Curated Images): Extensively pre-processed dataset featuring diverse poses, background isolation/removal on key subjects, extreme detail macro crops, and multiple real female athletes.
The "Extreme First" Philosophy: This LoRA was intentionally trained on peak/extreme female conditioning examples. Why? It is always easy to dial down a strong LoRA (e.g. to 0.4–0.6) to achieve a natural or superheroine look without losing texture quality. Conversely, trying to push a "weakly trained" LoRA past 1.0 quickly breaks image quality.
🔑 Main Trigger & Trained Keywords
Main Trigger Word: wellness_physique
Trained Vocabulary / Keywords: muscular, vascular, prominent vascularity, muscle striations
⚙️ Weight Settings
Use whatever weight works best for your prompt and checkpoint!
1.0 or lower: Dial the weight down if you want less of the LoRA's effect.
Above 1.0: Not recommended. Pushing the weight past 1.0 will cause image quality to suffer and ruin the render.
📌 Note on Showcase Images: All preview render examples shown on this page were generated at 1.0 LoRA strength unless explicitly specified otherwise.
🧪 Sampler & Generation Settings
Tested Sampler: euler (Used in author showcase)
Tested Scheduler: simple (Used in author showcase)
CFG: 1.0
Steps: 12 – 20+
💡 Experimentation Encouraged: These settings (euler / simple) are simply what I used during testing and for the preview renders. I haven't tested every possible combination, so please go ahead and experiment with different samplers and schedulers to see what works best for your specific goal (photorealism, stylized art, superheroine looks, etc.)!
Best Tested Checkpoints So Far:
Krea2 Turbo (FP8 scaled)
⚠️ Note: This model was trained/tested within the Krea2 ecosystem. Optimal settings are still actively being explored—feel free to post your render combinations and findings in the comments!
📝 Example Showcase Prompt
Official parody movie poster for "WELLNESS WOMAN". Dual composite poster layout showing two shots of the same hyper-muscular wellness_physique woman. On the left side, a quarterturn front pose; on the right side, a full backshot rear pose displaying her upper back, massive glutes, hamstrings, and calves with extreme vascularity and muscle striations. Large bold title text "WELLNESS WOMAN" centered at the top in metallic gold typography. She wears a dark athletic baseball cap with "100% NATTY" embroidered in white text, and holds a luxury handbag with a parody "LIFT VUITTON" logo. Pristine red and gold corset bodysuit, zero battle scars. Barefoot on a rocky beach by the sea under an overcast sky. RAW film photograph, natural skin pores, matte texture, clean graphic poster design, no plastic skin, no extra watermarks.Description
v3.0 - Realism Focused Training
New Training Approach: v3 is a fresh training attempt using an updated dataset and training setup aimed at pushing maximum muscle realism and lifelike detail.
Key Improvements: Enhanced muscle separation, quad sweep, glute definition, and support for realistic vascularity/striations.
Work in Progress: This version is not perfect yet. Further updates will address remaining minor issues to improve overall consistency. Feedback and user generations are greatly appreciated to help refine future versions!
FAQ
Comments (8)
Working 100% for me, thanks bro!
can we get a general 'muscle' slider lora? not so much into veins and low body fat stuff
Thanks for the feedback! That’s a great question, and it touches on something I actively experimented with during development.
My goal with capturing extreme, hyper-defined physiques was to prevent image quality degradation. As you might know, pushing LoRA weights high often causes color burn, artifacts, and degraded textures. My hypothesis was that training on extreme examples would allow users to stay in a low-weight 'safe zone' (around 0.2–0.5) to keep the overall image crisp and high-quality while still getting a noticeable muscle boost.
However, my testing showed that because the extreme dataset bundled muscle mass, vascularity, and low body fat together, scaling down the weight reduces all of those features at the same rate rather than stripping away just the veins.
While I work on a new training run specifically designed to isolate smooth muscle mass, here is how you can get that look with the current version without degrading image quality:
Negative Prompt: Add (veins, vascular, veiny, shredded, low body fat:1.2) to suppress the unwanted skin details.
Positive Prompt: Pair the trigger word with smooth skin, powerlifter physique, bulked.
Weight: Keep the weight low around 0.3 – 0.5 alongside the negative prompt above to maintain clean image quality.
@TestingAiImages I see, thank you
Works well for initial image generation. Tried it with Krea2Edit and unfortunately it doesn't seem to apply. Ended up using another female lora that added muscle via edit without issue.
Thanks for reaching out!
A very important detail when using this LoRA alongside krea2edit: make sure the Muscle LoRA is loaded before the krea2edit LoRA in your model chain / workflow (Base Model -> Muscle LoRA -> krea2edit LoRA).
Because krea2edit is an instruction/identity-editing adapter, loading it after the muscle LoRA ensures the anatomical features and muscle weights are applied to the base representation before the edit/restaging instructions take effect. If the order is reversed, krea2edit can suppress or override the muscle definition.
Try swapping the node/loading order and see if that fixes it! If you're still having issues, let me know your weight settings and node setup so we can troubleshoot further.
Cheers for the feedback. I set it up so the lora was before the edit lora and there was no change. I tried weights from 1 up to 3 and there is 0 difference to the output. When using the Muscular women lora (I simply put it into where the Wellness lora was), I can put in before/after the edit lora and it works without issue. The workflow I used was from here: https://huggingface.co/AiNinja94/Krea2-ComfyUI-Workflow/tree/main/Krea-2-Identity-edit
Also tested with:
Muscle / Fitness Slider (very high weight)= no change
Brute style = change
@jerbs Thanks for running those tests and sharing the detailed feedback—that’s really helpful context!
The reason broad LoRAs (like Brute style) apply easily while this one behaves like the slider is how Krea2Edit handles text conditioning. Unlike a style LoRA that forces a global image shift, this LoRA is trained on a precise anatomical and conditioning lexicon. In an identity-edit workflow, the model won't know to apply the LoRA’s weights unless you prompt for those specific body parts and conditioning tags.
Try incorporating the exact trained triggers directly into your edit prompt:
Main Trigger: wellness_physique
Conditioning & Detail: muscular, muscle striations, vascularity
Targeted Muscle Groups: muscular glutes, quads, hamstrings, calves, muscular lower body
Example edit prompt to test:
"photo of woman, wellness_physique, muscular, defined quads, muscular glutes, hamstrings, calves, muscle striations, vascularity"
A couple of other quick checks in ComfyUI:
CLIP & MODEL Routing: Make sure both the MODEL and CLIP wires pass through the Wellness LoRA node before reaching Krea2Edit. If CLIP is bypassed, the text encoder won't activate these specific tokens.
Denoise / Edit Strength: Because Krea2Edit is tuned heavily to protect the original body shape/identity, you may need to increase the edit strength/denoise slightly so it allows the muscle volume and definition to form.
Give those specific keywords a shot in the edit box and let me know if it starts picking up!