ntc-ai/krea2-particle-sliders

Krea2 Turbo-BBox Particle Sliders

Final Boss and Eldritch, with original rank-16 LoRAs and compressed rank-8 distills. Use strength 1 for the calibrated effect. Same prompt, seed and sampler across each comparison.

Built on jimmycarter/krea2-turbo-bbox, epoch-14-step-73184/transformer. These are ordinary attention LoRAs. The originals were trained directly as LoRAs; the distills compress those linear adapters.

Samples

Each comparison shows Off / On (Original), followed by Off / Distill. On and Distill both use strength 1. Click any image to open its full-resolution PNG. All images below are AI-generated, 768 × 768, 8 steps, guidance 0, mu=1.15. Each comparison states its seed and keeps it fixed across Original, Distill and Off. They were rendered from the released files, with no external alpha multiplier.

Final Boss

Robot in a rainy yard

A plain blue service robot gains angular shoulder and chest armor, heavier mechanical limbs and weathered panels. Same rainy factory yard, prompt and seed.

Seed 31415 · Side-by-side overview

Off	On (Original) · strength 1

Off	Distill · strength 1

Exact prompt

A full-body documentary photograph of an industrial robot standing outside after rain.
@35mm documentary photography, realistic wet metal, natural overcast daylight; Photograph
~A quiet factory yard with puddles, a brick wall and distant steel pipes.
o[230,80,770,950] A life-size humanoid service robot with a compact flat head, a single dark camera visor, plain faded blue metal panels and black mechanical joints, two arms and two legs, standing still with both hands lowered, feet firmly on the wet concrete.

Additional comparisons and unrelated fruit control

Heldout Bridge

Seed 42 · Side-by-side overview

Off	On (Original) · strength 1

Off	Distill · strength 1

Exact prompt

A lone warrior guarding a volcanic bridge.
@dramatic lighting, detailed fantasy game art; Digital illustration
~A stone bridge above a glowing lava river.
pe:1[220,80,780,950] An adult warrior in steel armor, holding a sword, standing guard.

Knight

Seed 42 · Side-by-side overview

Off	On (Original) · strength 1

Off	Distill · strength 1

Exact prompt

A full-body armored knight in a ruined cathedral.
@cinematic lighting, detailed game concept art; Digital illustration
~Ruined gothic arches and a cold stone floor.
pe:1[230,100,770,940] An adult knight in practical steel armor, plain helmet, holding a longsword, calm stance.

Street Photo

Seed 4242 · Side-by-side overview

Off	On (Original) · strength 1

Off	Distill · strength 1

Exact prompt

A candid full-body street photograph of a commuter on a rainy night in Tokyo.
@35mm street photography, realistic skin texture, natural proportions, cinematic neon reflections; Photograph
~A narrow city street with small restaurants, wet asphalt, red and blue neon reflections, soft background bokeh and gentle rain.
pe:1[240,100,760,950] An adult man with short dark hair in a simple dark wool overcoat, gray sweater, jeans and ordinary leather shoes, holding a closed black umbrella at his side, standing casually and looking toward the camera.

Fruit Control

Seed 42 · Side-by-side overview

Off	On (Original) · strength 1

Off	Distill · strength 1

Exact prompt

a bowl of fruit on a table

Workshop robot

Seed 2026 · Side-by-side overview

Off	On (Original) · strength 1

Off	Distill · strength 1

Exact prompt

A full-body photograph of a humanoid robot in an engineering workshop.
@industrial editorial photography, realistic metal and plastic, soft window light; Photograph
~A real workshop with concrete floors, workbenches and neatly arranged tools.
o[200,80,800,960] A life-size humanoid research robot with a simple rounded head, plain silver panels, exposed black joints and two ordinary arms, standing upright facing the camera.

Eldritch

Heldout Bridge

Seed 42 · Side-by-side overview

Off	On (Original) · strength 1

Off	Distill · strength 1

Exact prompt

A lone warrior guarding a volcanic bridge.
@dramatic lighting, detailed fantasy game art; Digital illustration
~A stone bridge above a glowing lava river.
pe:1[220,80,780,950] An adult warrior in steel armor, holding a sword, standing guard.

Additional comparisons and unrelated fruit control

Knight

Seed 42 · Side-by-side overview

Off	On (Original) · strength 1

Off	Distill · strength 1

Exact prompt

A full-body armored knight in a ruined cathedral.
@cinematic lighting, detailed game concept art; Digital illustration
~Ruined gothic arches and a cold stone floor.
pe:1[230,100,770,940] An adult knight in practical steel armor, plain helmet, holding a longsword, calm stance.

Fruit Control

Seed 42 · Side-by-side overview

Off	On (Original) · strength 1

Off	Distill · strength 1

Exact prompt

a bowl of fruit on a table

The bridge prompt was excluded from the original six-pair training set, then used for development comparisons. These examples are not a final-test benchmark. The featured Final Boss photograph was visually selected from eight robot photo prompts at strength 1, following an earlier eight-subject photo search. The Tokyo street photo and first workshop robot remain available in the additional examples. The robot selection notes and all eight comparisons and earlier photo search record the curation. The original Eldritch effect emphasizes organic armor and curling appendages; extra eyes and facial tentacles remain weak. The fruit control shows some rendering-style drift.

Downloads

Slider	Original · ComfyUI	Distill · ComfyUI	Original · Diffusers	Distill · Diffusers
Final Boss	Download	Download	Download	Download
Eldritch	Download	Download	Download	Download

Original files are about 77 MB; distills are about 38 MB. Both contain 128 projection adapters. PEFT originals · PEFT distills · Catalog · File hashes.

ComfyUI

Use standard Load LoRA with a ComfyUI export, MODEL strength 1, CLIP strength 0. No custom node is required. The optional comfy_krea2.py node, Krea2 Turbo-BBox LoRA under NTC/Krea2, also handles these files and the native exports' alpha metadata. Plugin ZIP · Setup.

Use krea2-bbox-turbo-comfy-latest.safetensors with the stock text encoder and VAE from Comfy-Org/Krea-2. Start from the Krea-2 Turbo template: 8 steps, CFG 1.0. The Diffusers equivalent is guidance_scale=0.0. The bbox checkpoint uses the distilled timestep shift mu=1.15.

Diffusers

Use a Diffusers build with Krea2Pipeline and Krea LoRA metadata support. The release validation records the tested versions. After loading the bbox transformer with the krea/Krea-2-Raw pipeline components:

pipe.register_to_config(is_distilled=True)  # selects mu=1.15
pipe.load_lora_weights(
    "ntc-ai/krea2-particle-sliders",
    weight_name="weights/native/krea2-eldritch-unit-alpha.safetensors",
    adapter_name="eldritch",
)
pipe.set_adapters("eldritch", adapter_weights=1.0)
image = pipe(prompt, height=768, width=768, num_inference_steps=8,
             guidance_scale=0.0).images[0]

The alpha is inside the file. For direct transformer.load_lora_adapter calls, use use_safetensors=True so the loader reads that metadata. PEFT users can load the corresponding folder and its adapter_config.json.

Grounded captions use x-first [x0,y0,x1,y1] boxes on a 0–1000 grid, one element per line, numeric character IDs, and panel-contained text. Prompting guide. The actual model tokenizer checked the 507-token content budget before training.

Distillation and alpha

Slider	Original rank / alpha	Distill rank / alpha	Projection relative MSE	Full-edit relative MSE	Edit cosine
Final Boss	16 / 16	8 / 8	0.000006	0.002129	0.998935
Eldritch	16 / 24	8 / 12	0.000006	0.001236	0.999382

Distills fit output principal components on 480 training activations per projection. Alpha selection uses 16 separate development states and complete denoiser edits. These errors measure approximation, not image quality. Compare the images before choosing a format. Method and all candidate measurements.

The originals retain every learned matrix. Final Boss embeds alpha 16 at rank 16; Eldritch embeds alpha 24 at rank 16, making released strength 1 equivalent to its previous strength 1.5. Student alphas are selected separately. Both formats keep the base's 8-step schedule; this distillation reduces adapter rank, not denoising steps.

Training and source

Both originals used physical GPU 0, rank 16, 400 updates, learning rate 5e-5, 512px, six paired captions, two cached trajectory seeds per pair, and preservation weight 0.1 every fifth update. The base and text encoder stayed frozen. Source, configurations, validation and reproduction belong to krea2-particle-sliders, following the release layout of anima-particle-sliders. The Krea originals are linear LoRAs; this is not the Anima nonlinear particle training recipe.

The GitHub repository does not ship slider weights or logs. Download weights from this Hub release. Reproduction · Training formulation · Release source archive · Source provenance.

python -m pip install -r requirements.txt
python scripts/train_krea2.py --dummy
python scripts/infer_krea2.py --help
bash scripts/train_final_boss_gpu0.sh
bash scripts/train_eldritch_gpu0.sh

The release tests real ComfyUI Krea modules, all 128 patch mappings, embedded alpha and clone isolation on CPU. Images are rendered through Diffusers. Full ComfyUI GPU generation is not part of this audit.

License

These adapters modify Krea 2 and are distributed under the Krea 2 Community License Agreement. By accessing or using these weights, recipients must agree to and be bound by that agreement. See NOTICE. This is an independent ntc-ai release, not an official or endorsed Krea product. Independently authored source code is MIT; that does not relicense the weights.