I trained a 1536px Krea 2 LoKr to test wardrobe & character consistency across 10 concert eras — here’s my full breakdown on LoKr vs. LoRA, epochs, and trigger tokens [Workflow Included]
retained text
Hey everyone! Over the past few weeks I fell down a rabbit hole comparing fine-tuning architectures for Krea 2 (trained against Krea 2 RAW, inferred on krea2\_turbo\_fp8\_scaled.safetensors). After training on RunPod, all of the checkpoint auditing, prompt engineering, and high-resolution generation was done locally in ComfyUI on my rig (RTX 5080 / Ryzen 9 9900X). I wanted to see how far I could push both facial consistency and complex couture stage wardrobe, so I set a goal to train a dedicated Krea 2 LoKr and stress-test it by recreating all 10 eras of The Eras Tour at native high resolution. Since the subreddit requires workflows, I’ve uploaded the exact ComfyUI `.json` workflow, the LoKr `.safetensors` weights, and a 40-image showcase gallery (4 frames per era) to Hugging Face (link at the bottom). Here is a breakdown of the training process, testing notes, and what worked best in case it helps anyone else experimenting with Krea 2 adapters. # 1. LoRA vs. LoKr Starting out, I was curious whether there was a noticeable difference between a standard LoRA and a **LoKr (Low-Rank Kronecker Product)** in Ostris's `ai-toolkit`. In a standard LoRA, the weight update matrix is formed by a simple inner product (`W = A × B`). LoKr uses the Kronecker product (`W = A ⊗ B`), which multiplies the ranks of the factor matrices together. The result is a compact \~29 MB file that packs a much higher effective rank and captures high-frequency spatial details—like natural skin micro-texture, individual sequin reflections, and subtle facial contours—without collapsing into that overly smooth, "plastic diffusion skin" look. I did also train a LoRA with the same parameters of the LoKr, but to my eye it looked worse at each comparable checkpoint which is why I ultimately settled on the LoKr. # 2. Curation of the 40-Image Dataset & Training on RunPod I curated a training dataset of 40 high-resolution reference photos, making sure to balance tight glam headshots, varied angles, and live stage lighting, each paired with its own `.txt` caption file. I wanted to train at a native **1536px resolution** to capture true pore-level skin fidelity and fine stage-costume micro-textures. Because 1536px training activations and Kronecker product gradient states push VRAM usage well past the 16GB ceiling on my local RTX 5080, I spun up a high-VRAM cloud GPU on RunPod to handle the heavy lifting. I ran the training through `ai-toolkit` out to 3,000 steps (75 epochs at batch size 1), saving checkpoints every 250 steps so I could pull them back down to my local 5080 rig and run a proper X/Y/Z epoch audit in ComfyUI: * **At 3,000 steps (75 epochs):** Because LoKr has such a high effective rank, it converges fast. By 75 epochs, the model was noticeably over-baked. The likeness was there, but the weights had started memorizing the exact lighting and rigid neck postures of the 40 training photos. When I tried prompting dynamic concert poses or heavy stage lighting, the adapter fought the prompt. * **At 1,500 steps (37.5 epochs):** The `00001500` checkpoint was the clear winner. Thirty-seven and a half passes per image hit the exact inflection point where facial geometry, eye spacing, and skin pores were locked in, while the weights stayed flexible enough to handle elaborate wardrobe changes and dynamic stage stances. # 3. The Trigger Word: "ts_person" vs. "Taylor Swift" During dataset prep, I captioned the 40 images using the synthetic trigger word `ts_person`. However, when I started testing checkpoints locally in ComfyUI, I ran a three-way comparison that completely changed how I prompt the model: 1. **Base Krea 2 +** `"Taylor Swift"` **(No LoKr):** Looks like a generic blonde celebrity lookalike or wax figure. Krea 2 knows who she is conceptually, but misses her exact facial geometry. 2. **Base Krea 2 + LoKr +** `"ts_person"` **Trigger:** Noticeably worse than expected but still better than no LoKr. Using an isolated synthetic token forces a 29 MB adapter to build a human identity from scratch in a vacuum, completely ignoring what the base model already knows. 3. **Base Krea 2 + LoKr (Model + CLIP Loader @ 1.0) +** `"Taylor Swift"`**:** Night and day difference. By routing both the Model and CLIP chains through the LoKr and prompting with her actual name (`Taylor Swift`), the adapter bridges the 40 training images directly onto Krea 2's pre-existing token understanding. The base model provides the macro structural context, and the LoKr acts as a precision correction layer that snaps her real facial features and skin fidelity into place. # 5. Prompting Couture Stage Wardrobe To stress-test the 1,500-step checkpoint locally on the 5080, I generated 40 showcase outputs (4 per era across all 10 eras, with 1 from each era featured in this post's gallery alongside the 2 covers). Instead of using generic tags like `sparkly bodysuit` (which causes models to hallucinate generic outfits or trigger modesty filters), using structural fashion terminology made the outfits pop. Not all outfits were easy to prompt for and some I straight up couldn't get to work even with LLM assistance. As an example of something that does work, here is the modular prompt chunk used for the **Lover** bodysuit: >`wearing iconic custom Atelier Versace tour costume: sleeveless performance bodysuit featuring a delicate pastel blush-pink and cerulean baby-blue ombré gradient, heavily encrusted with thousands of sparkling pave-set iridescent crystals catching stage beams, structured sweetheart bodice, high-cut leg opening, shimmering tan dance fishnet tights, matching custom knee-high crystal-embellished boots` # Local Hardware & ComfyUI Generation Specs * **Local Rig:** NVIDIA RTX 5080 / AMD Ryzen 9 9900X running ComfyUI * **Base Model:** `krea2_turbo_fp8_scaled.safetensors` (via `UNETLoader`, with LoKr trained against `krea2_raw`) * **Adapter:** `taylor_swift_krea_2_lokr.safetensors` (via `LoraLoader` Model + CLIP @ `1.0` / `1.0`) * **Sampler / Scheduler:** `euler` / `simple` * **Steps:** `8` * **CFG Scale:** `1.0` * **Resolutions:** `2880 x 2880` (Close-up Portraits) & `2496 x 3328` (Full-body Era shots) **Download Links (Weights, ComfyUI JSON Workflow & Full 40-Output Showcase Archive):** * **Hugging Face Repo:** [https://huggingface.co/BigGoober77/Taylor\_Swift\_Krea\_2\_LoKR](https://huggingface.co/BigGoober77/Taylor_Swift_Krea_2_LoKR) *(Note: The* `taylor_swift_krea_2_lokr.safetensors` *weights, the drag-and-drop* `krea2_Eras_Tour.json` *workflow, and all 40 full-resolution era renders are inside the* `Files` *tab).*
document fetched 2026-11-01 00:29 · canonical https://reddit.com/r/StableDiffusion/comments/1xkldof/i_trained_a_1536px_krea_2_lokr_to_test_wardrobe