deadman44/Flux_Photoreal_Models

Recent Updates

25/03/31 Add zipang_flux_ablit_test
24/12/9 Add zipang_flux_test02 (not good)
24/11/25 Add zipang_flux_test01

---

zipang flux ablit test

This model is experimental, taking the difference from flux1-dev and zipang_flux to LoRA and merging it into flux_ablit.
The quality is not good because it was necessary to increase the intensity of LoRA.
Its strength is x3.0 that of the original.

Download: zipang_flux_ablit_test-Q5_K_M.gguf (checkpoint)

19yo, myjd, A photograph of a young Japanese woman with straight dark hair styled in two braids, sitting on a stone pillar. She is wearing a white dress with a jacket. Her expression is happy, and she is looking directly at the camera. She has a slender build and fair skin. The background features a cityscape with tall buildings and a bridge, suggesting an urban setting. The lighting appears natural, with a soft, diffused quality, likely due to overcast skies. The camera angle is straight-on, providing a clear view of the girl and the background elements. The overall mood is serene and contemplative.
Steps: 12, Sampler: Euler, Schedule type: Simple, CFG scale: 1, Distilled CFG Scale: 3.5, Seed: 225447690, Size: 512x768, Model hash: b496e45346, Model: zipang_flux_ablit_test-Q5_K_M, Version: f2.0.1v1.10.1-previous-659-gc055f2d4, Diffusion in Low Bits: Automatic (fp16 LoRA), Module 1: ae_ablit, Module 2: clip_l_ablit, Module 3: t5xxl_bf16

---

zipang flux test01

-Finetune test model.
-Various age-group outputs
-flux_dev1 base

Download: zipang_flux_test01 (checkpoint)

-refer to png info

-VAE / Text Encoder:

ae, clip_l, t5xxl_fp16

-Sampling method

[Forge] Flux Realistic, Euler (Simple)

-Sampling steps

12-30

-CFG Scale

1

- sample prompt

17yo, myjk, grin, teeth,
Photograph of a Japanese girl with long dark hair in twintails, wearing a grey school uniform with a white shirt and a red plaid bow tie.
She is sitting at a desk in a classroom with a neutral expression, looking at the camera.
She is holding a pen.
The background shows a typical Japanese classroom with white walls.
The lighting is natural, likely coming from a window on the left side of the frame.
The camera angle is eye-level, capturing her from the front.
The photo has a casual, everyday feel.
Steps: 12, Sampler: Euler, Schedule type: Simple, CFG scale: 1, Distilled CFG Scale: 3.5, Seed: 1888130365, Size: 512x768, Model hash: db4f24fbb9, Model: zipang_flux_test01-Q5_K_M, Version: f2.0.1v1.10.1-previous-535-gb20cb4bf0, Diffusion in Low Bits: Automatic (fp16 LoRA), Module 1: clip_l, Module 2: t5xxl_fp16, Module 3: ae

- trigger

japanese, european, 
myjy, myjsl, myjsm, myjsh, myjc, myjk, myjd, 
and 3-30yo, 
and native english(recomended) or danbooru tags

---

zipang flux test02

-license: apache-2.0
-Base Model: Libre Flux
-not good...

Download: zipang_flux_test02 (checkpoint)

-VAE / Text Encoder:

ae, clip_l, t5xxl_fp16

-Sampling method

[Forge] Flux Realistic, Euler (Simple)

-Sampling steps

16-30

-CFG Scale

3

- sample prompt

20yo, myjd,
A photo of a Japanese girl standing on a street at night.
She has long black hair with bangs, and is wearing a light pink dress with lace details on the sleeves and has large-sized breasts.
She is looking directly at the camera with a sad expression, and her hands are behind her head.
The background is dark with some indistinct lights, likely from streetlights or a building, and neon light written "not good".
The lighting is artificial, with the photo taken at night.
The camera angle is straight-on, capturing her from the waist up.
The image is sharp with a shallow depth of field, focusing on the girl.
The photo has a high quality, professional feel.
Steps: 16, Sampler: [Forge] Flux Realistic, Schedule type: Simple, CFG scale: 3, Seed: 2680159760, Size: 512x768, Model hash: fab005c9da, Model: zipang_flux_test02-Q5_K_M, Version: f2.0.1v1.10.1-previous-535-gb20cb4bf0, Diffusion in Low Bits: Automatic (fp16 LoRA), Module 1: clip_l, Module 2: ae, Module 3: t5xxl_bf16

- trigger

japanese, european, 
myjy, myjsl, myjsm, myjsh, myjc, myjk, myjd, 
and 3-30yo, 
and native english(recomended) or danbooru tags

---

-Train Settings

• sd-scripts (SD3 branch)
  

base model: flux1-dev.safetensors, LibreFLUX.safetensors
vae/text encoder: clip_l.safetensors, t5xxl_fp16.safetensors, ae.safetensors

caption: JoyCaption Alpha Two
tags: WD EVA02-Large Tagger v3

--network_module "flux_train.py"
--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 "adafactor"
--optimizer_args "relative_step=False" "scale_parameter=False" "warmup_init=False"
--learning_rate=5e-7
--train_batch_size 7
--mixed_precision "bf16"
--save_precision "bf16"
--full_bf16
--loss_type "l2"
--huber_schedule "snr"
--timestep_sampling "flux_shift"
--model_prediction_type "raw"
--discrete_flow_shift 3.1582
--max_grad_norm=0
--min_snr_gamma=5
--apply_t5_attn_mask
--fused_backward_pass ^
--blocks_to_swap 35 ^
--skip_cache_check ^
--sdpa