<h1 id="anima-update">Anima Update</h1><p>Seems to be pretty good. If even just a functional background synthesizer, the model recognizes what the liminal images in the background are - while simultaneously holding the ground of character concentrated focus.</p><h2 id="uncertain-side-effects">Uncertain side effects</h2><p>Only slightly tested, but the dataset ran 20 epochs and the outcomes seem to be quite good for stylizing and simple prompting.</p><h2 id="simple-prompts-usable">Simple prompts usable</h2><p>The red haired girl with pink eyes wearing the blue dress in the liminal pool room seemed to generate just fine. I'll mess around with some more later.</p><h2 id="images-without-have-weak-backgrounds">Images Without Have Weak Backgrounds</h2><p>Generic prompts for the backgrounds tend to produce weaker effectiveness. Simultaneously with bigliminal, the backgrounds themselves become considerably more realistic for the anime characters. Not sure what other effects it has, but it definitely altered the model's perceptions of certain elements.</p><p>Match the prompt and you'll see the background, effectiveness, and solidity of the finetune fixed many of the background elements. Some of the elements that present themselves without the lora tend to be scattered, confusing, or disorienting; voxel overlap, disorienting shapes, and so on. With the liminal dataset it seems to have smoothed out the background in many elements for many room types.</p><p>A full background dataset wouldn't hurt.</p><h1 id="original-post">Original Post</h1><p>This is a subject oriented json experimentation meant to allow better control with json capacity to plain English. High probability that this variation faulted in one way or another. Likely some worse than others. It'll definitely produce some highly interesting and impressively unique liminal images due to the method trained.</p><h1 id="talking-to-the-models">Talking to the models</h1><p>Plain English is your best bet, or multilingual communication - if the model knows it. Use words as usual and then things will.. happen. Just ask the model what you want to see, and the model will create it for you. The more surreal, the more likely this model can accommodate than standard versions of this model. No guarantees.</p><h1 id="training">Training</h1><p>Each image was issued a json aligned caption organized by Qwen 3.5 0.8b finetuned to create json aligned structure.</p><h1 id="learned-what">Learned what?</h1><p>Json. The model learned how to think in Json. You don't have to understand why it does this, you simply have to understand the importance of subject association. In this case, the json links tokens together through subject association pathways.</p><p>Many of the pathways are interesting or useful, many of them corrupted or caused causal collapse. Which doesn't matter, because in a liminal training you WANT corruption and causal collapse.</p><h1 id="so-how-do-i-use-it">So how do I use it?</h1><p>Describe in plain text what you want to see. The model learned like this;</p><pre><code>{"subjects":[{"name":"housees","attributes":["yellow","rowed"]},{"name":"street","attributes":["asphalt"]},{"name":"grassy lanes","attributes":[]},{"name":"sidewalk","attributes":[]},{"name":"sky","attributes":["light"]}],"actions":["rowed in a line","street extends into the distance"],"setting":"outdoor"}</code></pre><p>What you say is;</p><p>A suburban neighborhood street extending into the distance, rows of houses along each sides with sidewalks and mailboxes per house.</p><h2 id="qwen-model">Qwen model</h2><p>This is the v2 lora trained for QWEN 3.5</p><p><a target="_blank" rel="ugc" href="https://huggingface.co/AbstractPhil/qwen3.5-0.8b-task_1-lora-v2">https://huggingface.co/AbstractPhil/qwen3.5-0.8b-task_1-lora-v2</a></p><p>Entire purpose is to convert english to json.</p><p></p><h2 id="larger-dataset-incoming">Larger dataset incoming</h2><p><a target="_blank" rel="ugc" href="https://huggingface.co/datasets/AbstractPhil/diffusion-pretrain-set-ft1">https://huggingface.co/datasets/AbstractPhil/diffusion-pretrain-set-ft1</a></p><p>A much larger pretrain dataset is preparing.</p><p></p><p></p>

BigLiminal - JSON Batch 1-2

2963438_training_data.zip

J7Y4MVAXD3SFYG8XN86EN6BKV0.safetensors

HiDream O1

Anima

Klein-9b

Klein-4b

Ernie

2963439_training_data.zip

PY2SYARRWMNZBA3Q9WHDQD3ZE0.safetensors

DK326HBD6Y5K9AS397GVF8SKA0.safetensors

2963436_training_data.zip

DWMMTS6056P9CWTW3SAP58H6N0.safetensors

2963437_training_data.zip

PWCXF9DNGXCZQ2B1Q70EA1JRW0.safetensors

2963419_training_data.zip

Files with SHA256: 54500746c4fa18c58292164e4fbc9d92f72123c315454b893b03f8bd580a197e