Another amazing model for creating dick or editing bad ones. It's getting hard to pick a favorite. Please share what you create via Add Post below.
This lora can be used with the base, turbo, or edit versions of Boogu Image. The 4 step turbo version does a decent job. For macro shots or extreme closeups, use the base model for more detail. You can modify images with the edit model to fix or add some D to your pics.
v1 can make it overly vascular if you mention veins, so prompt accordingly, or say smooth penis shaft and nothing about veins.
In general, remember with the turbo models that use cfg=1, any negative prompts will effectively be ignored (unless using NAG which currently doesn't exist for this model).
v1
Trained with my latest 200 image dataset. It trained for 7.5 hours for ~12k steps with AIToolkit using my 5090.
Description
12,277 steps for version 1 of this LoRA
FAQ
Comments (10)
thanks for your work.
Robot says:
For subject/concept LoRA training on this DiT architecture (Lumina/Boogu-based), you should exclude all embedding layers, conditioning pathways, and normalization parameters to avoid destabilizing the base model and interfering with timestep/conditioning handling.
Here are the substrings to exclude:
x_embedder ref_image_patch_embedder image_index_embedding time_caption_embed context_refiner noise_refiner ref_image_refiner norm_out norm1 norm2 norm3 ffn_norm1 ffn_norm2 attn_norm img_self_attn_normRationale:
Embedders (x_embedder, ref_image_patch_embedder, image_index_embedding): Input/output projections that shouldn't shift for concept learning
Conditioning blocks (context_refiner, noise_refiner, ref_image_refiner, time_caption_embed): Handle text/noise/reference image conditioning; training these causes overfitting and breaks base model behavior
Output norm (norm_out): Final layer normalization before output
AdaNorm/RMSNorm layers (norm1, norm2, norm3, ffn_norm1, ffn_norm2, attn_norm, img_self_attn_norm): These contain the timestep/conditioning injection parameters (.linear.weight, .linear.bias, .norm.weight) and layer-specific norm weights. Modifying these destabilizes training and affects how conditioning scales are applied to the residual streams.
The trainable parameters should be limited to the attention projections (to_q, to_k, to_v, to_out) and feedforward layers (feed_forward.linear_1/2/3) within single_stream_layers and double_stream_layers.
In ai-toolkit config yaml that looks like this. This is my 5th effort, looking good so far... network_kwargs: ignore_if_contains: - "x_embedder" - "ref_image_patch_embedder" - "image_index_embedding" - "time_caption_embed" - "context_refiner" - "noise_refiner" - "ref_image_refiner" - "norm_out" - "norm1" - "norm2" - "norm3" - "ffn_norm1" - "ffn_norm2" - "attn_norm" - "img_self_attn_norm"My vibe-coded lora loader does not produce good quality images with yours but does with mine. Smaller files, faster training, better results. ymmv
Not sure if you're a bot or what. Profile is pretty empty. I created a filtered version, and don't notice much difference. I'll put it under Optional Files and did one test which looks pretty much the same. Maybe the background is better text with the filtered version. See https://civitai.red/images/134910622
@coachbate Thanks for checking it out and training, appreciated. My profile is blank but I'm real, just anon. I'm checking out ComfyUI's native support I just found finally, will report anything useful.
@coachbate sorry for the ping. Some feedback since you went to the trouble of training in case it helps. Using ComfyUI native now. Very similar results with solo subject, once you introduce others it seems maybe the filtered model performs better with fewer incoherent outputs (detached/backwards penises etc). Limited gens so far. Apparently consistent with exclusion intent.
@civitaiclimatic796289 So sounds like I don't need to ask AIToolkit creator to incorporate this into his codebase for Boogu. I think Krea2 may have buried this model though.
@coachbate Your call for sure. I considered doing a PR for a sample config at least as a starting point. I think it's still best practice to target layers and exclude those that degrade, for any model, this paradigm would equally apply to Krea. I started looking at it yesterday...
@coachbate I ran the same prompt on the layer dump from Krea-2, but the result doesn't change with the excludes, still 256 modules. Here's the list in case it helps, not sure what to make of it yet. https://pastebin.com/0Bz7fk6x
(upd: reason is just how classes are coded, there exists "target_lora_modules" in the transformer source. Krea 2 out of the box isn't hitting 'bad' layers due to arch)
@civitaiclimatic796289 We were talking about boogu. please post on the Krea 2 model page I didn't filter anything on the krea 2. Let me know what your problem is on that page please- https://civitai.red/models/2727005 I'm getting great results with least number of throwaway bad output produced than ANY model yet.
@coachbate If you want better text or "dense" (?) text, Boogo's HF-page recommends to use Boogo Image Base for it https://huggingface.co/Boogu/Boogu-Image-0.1-Edit-Turbo#%F0%9F%94%AC-scenario-wise-comparison
@julesricardi373 I think most of my showcase is using base. I put it in the resources of each image