Shared core

Supra2-IMG trains the shared particle game. It does not keep a second copy of that game in this repository.

Pin particle-sliders-core from HyperGAN/particle-sliders:

particle-sliders-core @ git+https://github.com/HyperGAN/particle-sliders.git@a119ca1ecd3d5d6c437065839d22739b04f2f4d8#subdirectory=packages/particle-sliders-core

Import particle_sliders.winning_formulation. The pip name concept-slider-core and the import concept_slider_core are the old Anima extraction. This product does not install or import them.

from particle_sliders import winning_formulation

stamp = winning_formulation()
stamp.require({**stamp.as_dict(), "g_lr": 1e-4, "adv_batch": 4})
bridge = stamp.bridge()
regularizer = stamp.regularizer()
d_loss, g_loss, vic = stamp.losses()

require() locks the gmix architecture and the current formulation overlay (particle-gmix-1600-v2, provisional until ParticleGAN #38 crowns a full live leaderboard winner). Generator learning rate and batch size are a Supra model surface. Hub ids, prompts, and sampling are not formulation knobs.

Stays in particle-sliders-core	Stays in this repo
Routed particle adapter, global-mix critic, paired losses, particle VIC, gradient penalty	Train entry scripts/train_lora_supra.py
winning_formulation() and require()	Euler sample card: 256², 50 steps, CFG 3, t = i/K, z <- z + (1/K) * v
Formulation parameters	Base id SupraLabs/Supra2-IMG, product Hub id ntc-ai/supra-concept-sliders
	Prompt card configs/supra/prompts-supra.yaml
	Ordinary LoRA release on ntc-ai/supra-particle-sliders

vendor/particle-sliders/conceptmod/textsliders/supra_model.py is the DiT backbone used by ordinary LoRA inference. It is not the particle game. backend.lock.json records the core pin and that file's SHA-256. Do not point training at a PARTICLE_SLIDERS_ROOT checkout.

There is no ComfyUI node in this repository. When one exists, its class name stays here next to the Euler card.

Train

CPU smoke (no Hub weights):

python scripts/train_lora_supra.py --dummy --steps 8 --device cpu \
  --save-dir outputs/lighting-supra-dummy
python scripts/train_lora_supra.py --print-card

A CUDA run loads the pinned Supra2-IMG checkpoint and feeds mean-pooled Euler velocity edits into the same game:

python scripts/train_lora_supra.py --allow-hub --device cuda:0 \
  --save-dir outputs/lighting-supra

The published 1,600-step final-boss artifacts are ordinary LoRAs trained by scripts/train_final_boss.py. That recipe is separate from this game.

Research trainer still in particle-sliders

HyperGAN/particle-sliders still contains conceptmod/textsliders/train_lora_supra.py. That file is the research UNI trainer (FakeSupraBackend, trajectory / direct / cfg-delta losses in supra_slider.py). It does not call winning_formulation(), and its live path still stops before loading Flan-T5.

This pull request does not rewrite particle-sliders. Follow-up, in that repository only: replace train_lora_supra.py with a short comment pointing at scripts/train_lora_supra.py in HyperGAN/supra-particle-sliders. Leave supra_model.py there for the research dummy DiT; this repo already vendors that backbone for ordinary LoRA inference.