Krea-2 · Retro JP Magazine ‘88–‘95 — Full Fine-tune (Turbo / Raw / int8 / fp8 / nvfp4)

Works great with multiple photo collection pages.

I’ve wanted this for a long time: a Krea-2 model that gives you a late-80s / early-90s Japanese magazine page in one shot. Flash-lit gravure sets, film grain, multi-photo layouts on white paper, big headlines, covers and back covers.

This is a full fine-tune of Krea-2, not a LoRA. Getting here took a lot of scanning, tagging, filtering, layout captioning and many training runs. I really hope you enjoy it. ❤️

Versions-they need different steps and CFG for best quality

Turbo · bf16 (~25 GB): the fine-tune delta-merged onto the official Krea-2 Turbo. Best quality; 7 steps. CFG:1

Turbo · int8 ConvRot (~12.5 GB): recommended for most people. Closest to bf16, works on any GPU. Needs ComfyUI 0.27+.

Turbo · fp8 scaled (~12 GB): fast on RTX 40 / 50 series (Ada / Hopper / Blackwell).

Turbo · nvfp4 (~7 GB): only fast on RTX 50 / RTX PRO Blackwell. Smallest file, biggest quality loss.

Raw · bf16, full precision (~25 GB): the fine-tune itself, on Krea-2 Raw (not distilled). Use it for further training or your own merges. Needs about 30 steps at CFG 3.5.

You also need:

Text encoder qwen3vl_4b_fp8_scaled.safetensors (Qwen3-VL-4B). Load it with CLIPLoader, type krea2.

VAE qwen_image_vae.safetensors.

Put the model in models/diffusion_models and load it with Load Diffusion Model.

Activation words

No trigger word is needed. It’s a full fine-tune, so the style is baked in.

What steers it is the caption header every training caption started with: Magazine, Year, rating. All three parts are optional, but this order works best.

Magazine tags. Use the exact spelling if you perfer a magzine style

Deluxe Beppin (1988–1994)

Orange Tsu-Shin (1988–1994)

Apple Tsu-shin (1989–1995)

Actress Visual Movie Magazine (1991–1994)

Grace (1990–1994)

Sakuranbo Tsu-Shin (1991–1996)

URECCO (1991–1995)

Best Video (1995–2001)

BOndage

WOoooo Wolf (2004–2005): pulls toward a 2000s look

Weekly Playboy (2024): modern gravure, not retro

Year. Any year from 1988–1996 gives the strongest retro look; most of the data is 1991–1994. Later years exist but are sparse and look newer.

Rating word. Every caption had exactly one, and it strongly controls exposure:

clothed: swimwear, lingerie, fashion

topless

nude

explicit

Header examples:

Deluxe Beppin, 1992, clothed.

Orange Tsu-Shin, 1990, topless.

1993, clothed.

How the captions were written (copy this structure)

The model learned plain English sentences in three shapes.

Full magazine pages. Covers, back covers, full-bleed single-photo pages, multi-photo pages with 2–12 photos, and pages with article text. Example:

“A portrait magazine photo page with 3 photographs on a white background. The main photograph is a large landscape picture at the top of the page showing a three-quarter shot of a woman at eye level … The second photograph is a medium portrait picture below the main picture at the bottom-left showing … All 3 photographs show the same woman in the same outfit …”

Single photos. Example:

“A three-quarter photograph shot at eye level shows a woman sitting on … She wears … she has shoulder-length black hair … against a plain studio background with light from the front.”

Close-up / pose crops.

Multi-photo pages: prompt tips

Use a portrait canvas (2:3). For example 832×1248 or 1024×1536, or my two-pass setup below.

Start with the page type and the photo count: “A portrait magazine photo page with 4 photographs on a white background.”

Describe the main photograph first, in this order:

size: large / medium / small

orientation: landscape / portrait / square

position: top, bottom, top-left …

framing: full body / three-quarter / waist-up / close-up

angle: eye level / from above / from the side / from behind

what she is doing

Then each other photo in turn: “The second photograph is a medium portrait picture below the main picture at the bottom-left showing …”

Lock consistency with “All 4 photographs show the same woman in the same outfit …, shot in one session from different angles.” For variety, use “The photographs show different outfits or settings.”

Add typography last: “A large headline is printed at the top with two medium subheadings at the top-left and top-right, and a small block of body text at the bottom-right.” The printed text is decorative pseudo-Japanese, not readable.

2–5 photos is the sweet spot. 6–12 photos gives a contact-sheet look, but faces get tiny, so use the hires pass.

Works with your own character LoRA. Describe hair and face once, e.g. “In every photograph the woman has long black hair with blunt bangs.”

Example prompts

Multi-photo page:

Deluxe Beppin, 1992, clothed. A portrait magazine photo page with 3 photographs on a white background. The main photograph is a large landscape picture at the top of the page showing a full body shot of a woman at eye level lying on her side on a striped beach towel, propped on one elbow and smiling at the camera. The second photograph is a medium portrait picture below the main picture at the bottom-left showing a three-quarter shot of a woman at eye level sitting on the sand with her knees drawn up. The third photograph is a medium portrait picture below the main picture at the bottom-right showing a waist-up shot of a woman from a slightly low angle looking back over her shoulder. All 3 photographs show the same woman in the same outfit, a red high-cut one-piece swimsuit, on a sandy beach in bright daylight, shot in one session from different angles. A large headline is printed at the top-right.

Cover:

Grace, 1991, clothed. This portrait magazine cover has one photograph. The main photograph is a waist-up shot of a woman at a front angle wearing a white off-shoulder knit sweater, smiling at the camera against a pale blue studio background. A large headline is printed at the top with two medium subheadings at the top-left and top-right.

Single photo:

Orange Tsu-Shin, 1989, clothed. A three-quarter photograph shot at eye level shows a woman sitting on the edge of a hotel bed in a black lace slip dress, one hand resting on the sheets, looking at the camera, with on-camera flash and warm tungsten light in the background.

Workflow & settings (exactly what I use)

Turbo (bf16 / int8 / fp8 / nvfp4)

Load Diffusion Model → no ModelSamplingAuraFlow (shift off).

CLIPLoader: qwen3vl_4b_fp8_scaled, type krea2. VAE: qwen_image_vae.

Pass 1: 640×960, 7 steps, CFG 1.0, sampler er_sde, scheduler simple. euler / simple also works.

Pass 2 (hires): latent upscale ×1.9 (bislerp), then KSamplerAdvanced:

9 steps, start at step 4 (about 0.55 denoise)

CFG 1.0, euler / simple

final size ≈ 1216×1824

Single pass at 832×1248 is fine too.

At CFG 1 the negative prompt is ignored. Going above CFG ~1.5 burns the image.

Raw (bf16 Raw)

30 steps, CFG 3.5, euler / simple, and negative prompts work. Or use the official “Krea-2: Text to Image” Raw template: 25 steps, guidance 4.

About 4–5× slower than Turbo. Mainly meant for training and merging.

Training details

Base: full fine-tune of Krea-2 Raw with ai-toolkit.

Adafactor, lr 1.5e-5, effective batch 4, 768/1024 buckets

linear timesteps, caption dropout 0.1

10,750 steps

Data: 167 issues, mostly 1988–1996.

About 7,700 single photos, 3,300 full pages and 3,300 pose / detail crops

All four rating tiers, roughly balanced

Captions: two layers of vision-model tagging, fused into natural English by an LLM.

per photo: framing, angle, pose, outfit, hair, setting, light

per page: photo count, sizes, positions, text blocks

No names of any model or actress appear anywhere in the captions.

Turbo: delta merge, turbo_ft = turbo + α·(raw_ft − raw).

Quantized versions: made with convert_to_quant.

Please read: responsible use

18+ only. This model can produce nudity and explicit content. Rate your posts correctly.

Adults only. Every training page went through an age gate, and anything that could plausibly show someone under 18 was removed: school uniforms, school settings, childlike features. Don’t prompt for minors.

No real people. Don’t use this model to depict real, identifiable people.

Magazine names are style tags only. This model is not affiliated with or endorsed by any publisher.

Known limitations

Printed text is pseudo-Japanese.

In 6+ photo layouts faces get small; use the hires pass.

WOoooo Wolf and Weekly Playboy pull toward newer looks.

Scenes with two people are rarer and less reliable than solo gravure.

Have fun, and please share what you make! 📸