<h1 id="krea-2-retro-jp-magazine-'88-'95-full-fine-tune-(turbo-raw-int8-fp8-nvfp4)">Krea-2 · Retro JP Magazine ‘88–‘95 — Full Fine-tune (Turbo / Raw / int8 / fp8 / nvfp4)</h1><p></p><h2 id="works-great-with-multiple-photo-collection-pages.">Works great with multiple photo collection pages.</h2><p>Try in8 and nvfp4 if 24g is too big for you</p><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/660ae43c-6d4e-483d-90c2-4944c5f78ecb/original=true/660ae43c-6d4e-483d-90c2-4944c5f78ecb.jpeg" /></p><p>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.</p><p>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. ❤️</p><h2 id="versions-they-need-different-steps-and-cfg-for-best-quality">Versions-they need different steps and CFG for best quality</h2><p><strong>Turbo · bf16 (~25 GB):</strong> the fine-tune delta-merged onto the official Krea-2 Turbo. Best quality; 7 steps. CFG:1</p><p><strong>Turbo · int8 ConvRot (~12.5 GB):</strong> recommended for most people. Closest to bf16, works on any GPU. Needs ComfyUI 0.27+.</p><p><strong>Turbo · fp8 scaled (~12 GB):</strong> fast on RTX 40 / 50 series (Ada / Hopper / Blackwell).</p><p><strong>Turbo · nvfp4 (~7 GB):</strong> only fast on RTX 50 / RTX PRO Blackwell. Smallest file, biggest quality loss.</p><p><strong>Raw · bf16, full precision (~25 GB)</strong>: 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.</p><h3 id="you-also-need:">You also need:</h3><h3 id="text-encoder-qwen3vl_4b_fp8_scaled.safetensors-(qwen3-vl-4b).-load-it-with-cliploader-type-krea2.">Text encoder qwen3vl_4b_fp8_scaled.safetensors (Qwen3-VL-4B). Load it with CLIPLoader, type krea2.</h3><h3 id="vae-qwen_image_vae.safetensors.">VAE qwen_image_vae.safetensors.</h3><h3 id="put-the-model-in-modelsdiffusion_models-and-load-it-with-load-diffusion-model.">Put the model in models/diffusion_models and load it with Load Diffusion Model.</h3><h3 id="activation-words">Activation words</h3><p><span style="color:rgb(250, 82, 82)">No trigger word is needed. It’s a full fine-tune, so the style is baked in.</span></p><p>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.</p><p><span style="color:rgb(64, 192, 87)"><strong>Magazine tags.</strong></span> Use the exact spelling if you perfer a magzine style</p><p>Deluxe Beppin (1988–1994)</p><p>Orange Tsu-Shin (1988–1994)</p><p>Apple Tsu-shin (1989–1995)</p><p>Actress Visual Movie Magazine (1991–1994)</p><p>Grace (1990–1994)</p><p>Sakuranbo Tsu-Shin (1991–1996)</p><p>URECCO (1991–1995)</p><p>Best Video (1995–2001)</p><p>BOndage</p><p>WOoooo Wolf (2004–2005): pulls toward a 2000s look</p><p>Weekly Playboy (2024): modern gravure, not retro</p><p><span style="color:rgb(64, 192, 87)">Year.</span> 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.</p><p><span style="color:rgb(130, 201, 30)">Rating word.</span> Every caption had exactly one, and it strongly controls exposure:</p><p>clothed: swimwear, lingerie, fashion</p><p>topless</p><p>nude</p><p>explicit</p><p>Header examples:</p><p>Deluxe Beppin, 1992, clothed.</p><p>Orange Tsu-Shin, 1990, topless.</p><p>1993, clothed.</p><h1 id="how-the-captions-were-written-(copy-this-structure)"><span style="color:rgb(250, 82, 82)">How the captions were written (copy this structure)</span></h1><p>The model learned plain English sentences in three shapes.</p><p>Full magazine pages. Covers, back covers, full-bleed single-photo pages, multi-photo pages with 2–12 photos, and pages with article text. Example:</p><p>“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 …”</p><p>Single photos. Example:</p><p>“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.”</p><p>Close-up / pose crops.</p><p>Multi-photo pages: prompt tips</p><p>Use a portrait canvas (2:3). For example 832×1248 or 1024×1536, or my two-pass setup below.</p><p>Start with the page type and the photo count: “A portrait magazine photo page with 4 photographs on a white background.”</p><p>Describe the main photograph first, in this order:</p><p>size: large / medium / small</p><p>orientation: landscape / portrait / square</p><p>position: top, bottom, top-left …</p><p>framing: full body / three-quarter / waist-up / close-up</p><p>angle: eye level / from above / from the side / from behind</p><p>what she is doing</p><p>Then each other photo in turn: “The second photograph is a medium portrait picture below the main picture at the bottom-left showing …”</p><p>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.”</p><p>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.</p><p>2–5 photos is the sweet spot. 6–12 photos gives a contact-sheet look, but faces get tiny, so use the hires pass.</p><p>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.”</p><p>Example prompts</p><h3 id="multi-photo-page:">Multi-photo page:</h3><p>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.</p><h3 id="cover:">Cover:</h3><p>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.</p><h3 id="single-photo:">Single photo:</h3><p>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.</p><h2 id="workflow-and-settings-(exactly-what-i-use)"><strong>Workflow &amp; settings (exactly what I use)</strong></h2><h3 id="turbo-(bf16-int8-fp8-nvfp4)">Turbo (bf16 / int8 / fp8 / nvfp4)</h3><p>Load Diffusion Model → no ModelSamplingAuraFlow (shift off).</p><p>CLIPLoader: qwen3vl_4b_fp8_scaled, type krea2. VAE: qwen_image_vae.</p><p>Pass 1: 640×960, 7 steps, CFG 1.0, sampler er_sde, scheduler simple. euler / simple also works.</p><p>Pass 2 (hires): latent upscale ×1.9 (bislerp), then KSamplerAdvanced:</p><p>9 steps, start at step 4 (about 0.55 denoise)</p><p>CFG 1.0, euler / simple</p><p>final size ≈ 1216×1824</p><p>Single pass at 832×1248 is fine too.</p><p>At CFG 1 the negative prompt is ignored. Going above CFG ~1.5 burns the image.</p><h2 id="raw-(bf16-raw)">Raw (bf16 Raw)</h2><p>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.</p><p>About 4–5× slower than Turbo. Mainly meant for training and merging.</p><p>Training details</p><p>Base: full fine-tune of Krea-2 Raw with ai-toolkit.</p><p>Adafactor, lr 1.5e-5, effective batch 4, 768/1024 buckets</p><p>linear timesteps, caption dropout 0.1</p><p>10,750 steps</p><p>Data: 167 issues, mostly 1988–1996.</p><p>About 7,700 single photos, 3,300 full pages and 3,300 pose / detail crops</p><p>All four rating tiers, roughly balanced</p><p>Captions: two layers of vision-model tagging, fused into natural English by an LLM.</p><p>per photo: framing, angle, pose, outfit, hair, setting, light</p><p>per page: photo count, sizes, positions, text blocks</p><p>No names of any model or actress appear anywhere in the captions.</p><p>Turbo: delta merge, turbo_ft = turbo + α·(raw_ft − raw).</p><p>Quantized versions: made with convert_to_quant.</p><p>Please read: responsible use</p><h2 id="18+-only.-this-model-can-produce-nudity-and-explicit-content.-rate-your-posts-correctly.">18+ only. This model can produce nudity and explicit content. Rate your posts correctly.</h2><h2 id="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.">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.</h2><h2 id="no-real-people.-don't-use-this-model-to-depict-real-identifiable-people.">No real people. Don’t use this model to depict real, identifiable people.</h2><h2 id="magazine-names-are-style-tags-only.-this-model-is-not-affiliated-with-or-endorsed-by-any-publisher.">Magazine names are style tags only. This model is not affiliated with or endorsed by any publisher.</h2><h2 id="known-limitations">Known limitations</h2><h2 id="printed-text-is-pseudo-japanese.">Printed text is pseudo-Japanese.</h2><h2 id="in-6+-photo-layouts-faces-get-small-use-the-hires-pass.">In 6+ photo layouts faces get small; use the hires pass.</h2><h2 id="wooooo-wolf-and-weekly-playboy-pull-toward-newer-looks.">WOoooo Wolf and Weekly Playboy pull toward newer looks.</h2><h2 id="scenes-with-two-people-are-rarer-and-less-reliable-than-solo-gravure.">Scenes with two people are rarer and less reliable than solo gravure.</h2><h2 id="have-fun-and-please-share-what-you-make!">Have fun, and please share what you make! 📸</h2><p></p>

Files with SHA256: fb3fc127c9ca4c77ad2e264fe769b617f45df1eda60406745ffe3bfaaf1326d3