metadata

license: other
license_name: flux-1-dev-non-commercial-license
license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md
language:
  - en
tags:
  - text-to-image
  - stable-diffusion
  - safetensors
  - stable-diffusion-xl

myjsm_flux_test02

-It is a test lora of poor quality with only a few images learned.
-trained 273 images.

Found a slightly better training setting. But still hard to find things that don't show up in flux.

Download:test02

The model used for the test is Flux Fusion DS v0 GGUF Q4_0 (UNET) and v0 GGUF Q5_K_M.

GGUF Q4_0 + t5xxl_fp8_e4m3fn : 4step

GGUF Q5_K_M. + t5-v1_1-xxl-encoder-Q5_K_M : 12step

-refer to png info

- sample prompt

9yo, myjsm, japanese,
photograph of a girl sitting on a brick pavement with a pink umbrella in front of her. She is wearing a white camisole and a blue skirt with a anime print. She has shoulder-length dark hair and is smiling at the camera.
bangs, black eyes, skirt, rain
<lora:myjsm_flux_test02:1>
Steps: 4, Sampler: Euler, Schedule type: Simple, CFG scale: 1, Seed: 642879771, Size: 792x1056, Model hash: 5e21feb505, Model: FusionDS_v0_Q4, Lora hashes: "myjsm_flux_test02: 3fdff20b7d65", Version: f2.0.1v1.10.1-previous-419-gf82029c5c, Module 1: ae, Module 2: clip_l, Module 3: t5xxl_fp8_e4m3fn

- trigger

myjsm, japanese, 9yo, 
and native english

-Train Settings

base model: flux1-dev.safetensors
vae/text encoder: clip_l.safetensors, t5xxl_fp8_e4m3fn.safetensors, ae.safetensors
tag: caption (native eng) + tags (danbooru)

--network_module "networks.lora_flux"
--gradient_checkpointing
--cache_latents
--cache_latents_to_disk
--cache_text_encoder_outputs
--cache_text_encoder_outputs_to_disk
--enable_bucket
--bucket_no_upscale
--optimizer_type "AdamW8bit"
--optimizer_args "weight_decay=0.01" "betas=0.9,0.999"
--learning_rate=0.0002
--network_dim=32
--network_alpha=4
--network_train_unet_only
--mixed_precision "bf16"
--save_precision "bf16"
--full_bf16
--loss_type "l2"
--huber_schedule "snr"
--model_prediction_type "raw"
--discrete_flow_shift 3
--timestep_sampling "sigma"
--max_grad_norm=1
--max_timestep=1000
--min_snr_gamma=5
--min_timestep=100
--noise_offset=0.0375
--adaptive_noise_scale=0.00375
--apply_t5_attn_mask
--split_mode
--network_args "loraplus_unet_lr_ratio=16" "train_blocks=single"

myjsm_flux_test01

-It is a test lora of poor quality with only a few images learned.
-trained 167 images.

I don't know why the image turns white.

Download:test01

The model used for the test is Flux Fusion DS v0 GGUF Q4_0 (UNET)

test01	original

myjsm, japanese, photograph of a portrait of a young girl sitting on a white couch in girly room. She is wearing a pink and white checkered dress with a pink bow on her head. black hair, twintails, bangs, black eyes, smile, upper body

- trigger

myjsm, japanese, 9yo

-Train Settings

base model: flux1-dev.safetensors
vae/text encoder: clip_l.safetensors, t5xxl_fp8_e4m3fn.safetensors, ae.safetensors
tag: caption (native eng) + tags (danbooru)

--network_module "networks.lora_flux"
--gradient_checkpointing
--cache_latents
--cache_latents_to_disk
--cache_text_encoder_outputs
--cache_text_encoder_outputs_to_disk
--enable_bucket
--bucket_no_upscale
--optimizer_type "AdamW8bit"
--optimizer_args "weight_decay=0.01" "betas=0.9,0.999"
--learning_rate=0.0002
--network_dim=32
--network_alpha=16
--network_train_unet_only
--mixed_precision "bf16"
--save_precision "bf16"
--full_bf16
--loss_type "l2"
--huber_schedule "snr"
--model_prediction_type "raw"
--discrete_flow_shift 3
--timestep_sampling "sigma"
--max_grad_norm=1
--max_timestep=1000
--min_snr_gamma=5
--min_timestep=100
--noise_offset=0.0375
--adaptive_noise_scale=0.00375
--apply_t5_attn_mask
--split_mode
--network_args "loraplus_unet_lr_ratio=16" "train_blocks=single"