V2 Update: Pruned version, plus refinements. Both full size and pruned v2.0 models are available. Features:

• Slightly increased realism of skin textures and lighting

• Better adherence to clothing prompts

• Increased stylized facial details

DTF (deuterium-tritium fusion) Mix is a high fidelity 2.5D/3D model aimed at generating complex lighting, clothing, characters and environments. It accomplishes this by having concepts from a wide range of real, anime, cartoon, illustration and (of course) pony data. The mix combines capabilities of more realism-oriented models with the aesthetic effects and character designs seen in 2.5D and hyperrealistic styles. DTFMix excels at illustrating fantasy and sci fi settings, NSFW situations and detailed anatomy.

This is a one-shot model. Due to its high success rate with anatomy and interactions, the model can one-shot effectively up to around 1440x1440 resolution. However, I always suggest using one pass of ADetailer with faceyolov8n with any PonyXL model.

/Fixed with v2.0/ Why is this model 11.4gb? The pruned version of this model has noticeably less versatility than the full version. The only drawback of this larger model size is that some GPUs will not be able to use the model at higher resolutions.

DTFMix Use Guide

Here are basic starting parameters for users plus more advanced prompts.

DTFMix is intended to one-shot with minimal user effort, meaning the output does not need to be upscaled or inpainted when genning at intermediate resolutions (though an ADetailer pass is recommended)

Parameters

• Tested at 3-7cfg, with ~5cfg being the suggested start point.

• I typically use Euler A at 27 steps, though lower steps (17-23 steps) can also produce good results.

• Resolutions between 1280x940 / 940x1280 and 1400x1600 / 1600x1400 have been tested.

Style

DTFMix can be tuned to produce a variety of styles. Most Pony style tags are supported: 2D, illustration, sketch, lineart, traditional media, painted, 3D, realism, realistic, source anime, source photo, flat shading, cel shading, traditional media,

Quality+Effects

• Supported quality prompts include all terms normally seen in Pony: score_9, score_8_up, score_7_up, best quality, high quality, etc.

• There are also quality prompts that will tend to impact overall style, useful for detailed environments with less prompting: ultra-detailed, intricate details, very aesthetic, 8k, 16k,

• DTFMix responds well to optical effects such as: chromatic aberration, subsurface scattering, lens distortion, sparkling, film burn, wide angle lens, godrays, rim lighting, depth of field, dramatic lighting, volumetric lighting, reflections,

Composition

Character-focused images can be prompted for a wide range of perspectives: closeup, dutch angle, pov, dramatic angle, distinct poses, distinct actions, from below, extreme angle, selfie, foreshortening,

LORA Compatibility

The model has good resistance to "frying" from overtrained LORAs while also easily adopting a true 2D/lineart style when LORAs are used. Using a LORA, you can quickly force a true 2D style.

Verbose Prompting

DTFMix responds surprisingly well to verbosity. Try to limit verbose prompts to 20-30 tokens maximum.