<h1 id="supports-up-to-512-tokens-with-alucard">Supports up to 512 tokens with Alucard</h1><h1 id="what-happens-when-it-goes-badly">What happens when it goes badly?</h1>Hyperfixation. It will likely choose a topic and hyperfixate on it. At some point the bias ended up defaulting some of those elements too heavy, so the system needs a full reweight for a proper utilization.<h1 id="how-to-use">How to use?</h1>The ComfyUI workflow is attached to the businesswomen pictures to make Alucard work how it's intended, but it's not the simplest process.Alright I correctly named everything now. So if you got the wrong one ensure you get the correct version.<h1 id="prompting">Prompting</h1>Each version is trained to activate based on the size 1280x1280 or larger up to around 16nnx16nn and the NSFW elements should be flat blacklisted.Supports up to 512 tokens with alucard and the flan-t5-large or xl set to 512 tokens. A bit more stable at 256 but it can definitely handle it. Pairing it with bert-beatrix-2048 allows an additional context control BEYOND 512 up to 2048 but the outcome from this was very shaky at best.They are predominantly pony inputs for quality; so you can use<ul><li>score_9, score_8, score_7, score_6, score_5, score_4, score_3, score_2, score_1 for score</li><li>masterpiece, most aesthetic, very aesthetic, displeasing, and very displeasing for aesthetic</li><li>rating_safe, rating_sensitive, rating_explicit, rating_nsfw</li><li>grid_a1 &lt;- full grid support</li><li>plain english support with many many elemental systems at play.</li><li>a large array of color, quality, and deterministic capacities.</li></ul><h1 id="making-it-work">Making it work</h1>SIMULACRUM ROSE Epsilon has three components:<ol><li>Refiner - specifically meant to refine steps 800-1000 aka 80% to 100% of the image generation.<ol><li>Requires:<ol><li>Simulacrum V4 - Epred expert + the clips</li></ol></li></ol></li><li>Early - specifically meant to context solidify and completely gut images for text to image generation.<ol><li>Requires:<ol><li>Simulacrum V4 - Epred expert + the clips</li></ol></li></ol></li><li>Full - a combined effort between the two after merging and full finetuning the system 0-1000.<ol><li>Requires:<ul><li>Simulacrum V4 - Epred expert + the clips</li><li><a target="_blank" rel="ugc" href="https://github.com/AbstractEyes/comfy-clip-shunts/tree/dev2_electric_boogaloo">Alucard</a><ul><li>dev2_electric_boogaloo</li><li>clip sampler configured<ul><li>there are multiple currently, use the configured</li></ul></li><li>shiva folding</li><li>100 steps 1 pass<ul><li>flan-t5-large</li><li>bert-beatrix-2048</li></ul></li></ul></li><li>Simulacrum Rose T2I - 1.0 strength</li><li>All of that to make the full function correctly. They might be mislabeled, so swap refiner for t2i if you get scanlines or odd results from your output.</li></ul></li></ol></li></ol><h1 id="training-process">Training process</h1>Each trained with about the same 1 million images for over 100 epochs, the problems start to manifest quickly. They are very overcooked which is why I've been reluctant to release them.There is NO EASY WAY to get the proper interpolated implementation of Simuacrum Rose - Full without a big toolkit surrounding it. David will solve this problem and pair directly with Simulacrum Rose V2 in the future.They were each trained using a new format of scheduling, folding, and processing called Alucard - which is a stepping process meant to create compartmentalized bucketing of conditioning alongside the standard noise generation.The system failed, essentially, because it could not converge with the noise types with the current spectrum of interpolation. This is partly due to the differential equations being incorrect and partly due to the drift introduced BY the stepping system being randomThese are essentially NSFW models because they burned so badly, but they are still utilizable for many utilities and are quite potent for utilization.<h1 id="it-definitely-didn't-turn-out-as-good-as-expected.">It definitely didn't turn out as good as expected.</h1>I deeply apologize for those waiting for this one, as I had put it off for quite a while, and the results turned out to be incompatible with the spectrum. Yet the output should be shared nonetheless.The token context will retain cohesion with many tokens beyond expectation. It's been trained with bert-beatrix-2048 and flan-t5-large as catalysts so it functions best with those when used with the standard clips paired with simulacrum-epsilon-v4.

SDXL Simulacrum ROSE 512 tokens T5-XL

This is the 0-800 version, meant specifically to generate text to image.

simulacrum-rose-full-e100.safetensors

simulacrum-rose-early-e100.safetensors

Simulacrum Rose Full Workflow.zip

Full - 0-1000

T2I - 0-800

Refiner - 800-1000

This is the full variation. They are all overcooked, but they are still interesting.

Files with SHA256: 168da6a3184b7847052addd0c6b5bde44e460c91acc410cba535755a65b4503d