Experimental test run — trained and auto-captioned entirely using Civitai Trainer. No local training, no manual captions. Proof-of-concept, not a polished release. May not work without careful prompting. (try forcing to use Chromostereopsis style with red blue and black colors without any glow or gradients) Produces a chromostereopsis style — vivid red subjects on deep blue backgrounds, creating a pulsating depth illusion. Dataset was generated by a personal SDXL LoKR from 2023 (available on request).<h2 id="training-details">Training Details</h2><ul><li>Base model: Flux.2 Klein 9B</li><li>Engine: ai-toolkit (Civitai Trainer) — training + auto-captioning done entirely on-platform</li><li>Network: LoRA rank 64 / alpha 32</li><li>Epochs: 15 | Repeats: 20</li><li>Resolution: 1024</li><li>Optimizer: AdamW 8-bit | LR: 1e-4 | Scheduler: cosine</li><li>Text encoder: not trained</li></ul>The training dataset was generated by a personal SDXL LoKR from 2023 that established the core red-blue palette. Happy to share the dataset and the source LoKR if anyone wants to experiment further.Usage: Mention the red/blue color scheme in your prompt. Weight 0.7–1.0. Expect inconsistency.

Chromostereopsis

<pre><code>{
 "lr": 0.0001,
 "engine": "ai-toolkit",
 "epochs": 15,
 "ecosystem": "flux2klein",
 "keepTokens": 0,
 "networkDim": 64,
 "numRepeats": 20,
 "resolution": 1024,
 "lrScheduler": "cosine",
 "minSnrGamma": null,
 "noiseOffset": null,
 "modelVariant": "9b",
 "networkAlpha": 32,
 "optimizerType": "adamw8bit",
 "shuffleTokens": false,
 "textEncoderLr": null,
 "flipAugmentation": false,
 "trainTextEncoder": false
}</code></pre>

Files with SHA256: 66757a4174b2d75af4bb59e5bc566e7ea10707396aae4107bda64f9b7e60cc57