━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✨ **Krea2 Turbo — QwenVL All-in-One (Auto-VRAM Adaptive)**
ComfyUI · Krea-2 Community License · Self-Adapting VRAM Engine + Measured Peak-VRAM Proof
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

The complete Krea2 Turbo dual-mode workflow — detail rebalance, inline SeedVR2 upscale, and NSFW-tuned prompting — merged into one graph, plus a new headline feature: **the pipeline reads your GPU's total VRAM at queue time and configures itself automatically.** No manual "low VRAM / high VRAM" switch to pick by hand. Every run also prints its own **measured peak VRAM** to console — not a marketing claim, an actual number from your card, your run.

Same Krea2 Turbo FP8 base as all prior versions on this page (no LoRA, no base model swap, same euler/simple/CFG1.0/10-step sampler). Tested on RTX 5080 16GB.

⚠️ **Read the Honest Disclosure section before downloading** — the auto-VRAM claim is scoped to fresh-queue generation. Please read it before assuming "under 8GB" holds for every scenario.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🆕 **What's New — the auto-VRAM-adaptive engine**

This is the differentiator: three cooperating VRAM mechanisms, all new, none manual.

🎯 **Self-Adapting VRAM Cap (always on)** — a headroom-control node caps the core pipeline's peak usage at a target you can tune (default 6.5GB headroom target). Full nvfp4/fp8 quality — this is not a quantization swap, it's smarter model eviction timing. Measured peak with the default target: **~6.8GB**, repeatable across runs.

🔀 **Upscale Auto-Switch** — reads your GPU's total VRAM once at queue time. ≥10GB total → runs the inline SeedVR2 upscale stage automatically. <10GB total → skips it and stays at native resolution. No switch to flip by hand; manual override still available if you want to force one way or the other.

🔀 **Quality-Tier Auto-Switch** — a second auto-decision, same 10GB threshold, that sets the upscale stage's own internal resolution target and I/O-swap behavior together: **Low tier** (below threshold) runs a smaller 1536px upscale target with component swap enabled, measured fresh-queue peak **~7.0–7.4GB**. **Full tier** (at/above threshold) runs the full 2176px target with swap disabled, measured fresh-queue peak **~8.6–9.0GB**.

📟 **Peak-VRAM Proof Stamp** — every run prints its own measured peak (both allocated and reserved CUDA memory, not a single cherry-picked number) to console. See your own card's real number, not our claim.

**Unchanged from prior versions on this page:**
✅ Same Krea2 Turbo FP8 base · same euler / simple / CFG 1.0 / 10-step sampler
✅ Same dual-mode (keyword / reference) + dual-orientation switch
✅ Detail Rebalance toggle (new, this version) · NSFW-tuned prompt toggle (from v2) — both still user-controlled, default On / Off respectively
✅ No LoRA · no base model swap

**Version history:**
• v1 — SFW dual-mode QwenVL workflow (2026-06-28)
• v2 — NSFW + inline SeedVR2 (2026-07-05)
• v3 — All-in-one merge + Detail Rebalance + auto-VRAM-adaptive engine + peak-VRAM proof (this version)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔬 **Honest Disclosure — read this first**

The auto-VRAM claim is real but scoped. Here's exactly what was measured and what wasn't:

• **Fresh-queue numbers are reliable.** The first generation after opening ComfyUI (or after any idle period) consistently lands at the numbers quoted above — VRAM Cap ~6.8GB, Low-tier upscale ~7.0–7.4GB, Full-tier upscale ~8.6–9.0GB. These were measured directly via `nvidia-smi`, repeated, and cross-checked against actual output image resolution (not just the VRAM number) to confirm the auto-switch picked the branch it claimed to.

• **Back-to-back queued generations on the same long-running ComfyUI process can climb higher** — we observed peaks rising toward 10–14GB across consecutive runs in testing, even with explicit memory-free calls between them. This is a known PyTorch CUDA allocator behavior (memory fragmentation from repeated large-model swaps within one process), not a bug in the auto-switch logic — we confirmed the switch was still picking the correct branch every time (verified via actual output pixel dimensions), the memory just wasn't fully returning between runs.

• **Practical takeaway:** for the most predictable low-VRAM behavior in an extended session, restart ComfyUI periodically rather than queuing dozens of generations back-to-back without a break. A single fresh generation reliably hits the numbers above; a long unattended batch queue may not.

• **Why ship it anyway:** no competing Krea2 workflow we found auto-detects VRAM at all — the closest prior art is an unattributed boolean-only building block, not wired into any shipped workflow, and at least one major competing listing explicitly states no automatic VRAM detection exists in their design. This is a real, working, first-of-its-kind mechanism for this model family — just don't read "auto-adaptive" as "immune to how CUDA's memory allocator behaves over a long session."

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔞 **Mature Content Disclosure**

This version includes the same NSFW-tuned prompt-enhancer toggle introduced in v2 — off by default, switchable on. When enabled, the QwenVL prompt-enhancer system prompts support adult glamour, intimate, and nude photography output alongside general-purpose prompting.

All example showcase images depict fictional adult (18+) subjects generated entirely by AI — no real people, no likeness of any real individual. Individual showcase images are marked with their own explicit content rating (X/Mature) per Civitai's per-image rating system rather than flagging this entire listing page as mature.

Users are responsible for following Civitai's Terms of Service and their local jurisdiction's laws regarding generated adult content when using this workflow to create their own images.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✨ **Features**

✅ **Mode A: Keyword → Auto-Expand** — Type subject → QwenVL PromptEnhancer expands to rich visual prompt → generate
✅ **Mode B: Reference Image → Style Capture** — Drop reference image → QwenVL describes style & composition → generates inspired image
✅ **Dual Orientation** — landscape (16:9) or portrait (9:16), single switch, no node rewiring
✅ **Detail Rebalance Toggle** — optional conditioning-band boost for finer skin/fabric/hair texture (default On, ~0 extra VRAM cost)
✅ **NSFW Prompt Toggle** — optional adult-content prompt-enhancer wording (default Off)
✅ **Inline SeedVR2 Upscale, Auto-Switched** — runs automatically when your GPU has headroom, skips automatically when it doesn't
✅ **Self-Adapting VRAM Cap** — always-on headroom control, tunable target, full quality (no quantization tradeoff)
✅ **Peak-VRAM Proof Stamp** — measured (not claimed) peak VRAM printed every run
✅ **Krea-2 Original Architecture** — NOT FLUX-derived; DiT 12.9B model from Krea; FP8 quant by AlperKTS (~7GB)
✅ **Stable Sampler Config** — Euler, simple scheduler, CFG 1.0, 10 steps

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📦 **Required Models**

• krea2_turbo_fp8.safetensors (~7GB) — Main Krea-2 DiT diffusion model (FP8 quantized)
• qwen3vl_4b_fp8_scaled.safetensors (~2–3GB) — Combined text + vision encoder (Qwen3-VL-4B, FP8)
• qwen_image_vae.safetensors (~1GB) — Image VAE codec
• Qwen3-VL-2B-Instruct (~2.5GB, auto-downloads) — Vision model for prompt enhancement and image analysis
• seedvr2_ema_7b_fp16.safetensors — SeedVR2 7B DiT upscaler model (used when Upscale Auto-Switch enables the branch)
• ema_vae_fp16.safetensors — SeedVR2 VAE (tiled encode/decode)

⬇️ **Download**
• Krea2 base 3 files: https://huggingface.co/AlperKTS/Krea2_FP8 (unet / clip / vae folders)
• SeedVR2 2 files: https://huggingface.co/ByteDance-Seed/SeedVR2-7B (Apache-2.0) or fp8-variant mirror https://huggingface.co/numz/SeedVR2_comfyUI if VRAM-constrained

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🧩 **Required Custom Nodes**

1. **ComfyUI-QwenVL** (AILab / 1038lab) — PromptEnhancer + VL image analysis
2. **ComfyUI-Easy-Use** (vjumpkung) — anythingIndexSwitch for mode/orientation/content toggles
3. **ComfyUI-SeedVR2_VideoUpscaler** (numz) — inline hi-res upscale
4. **ComfyUI-Conditioning-Rebalance** (nova452) — detail-rebalance node — https://github.com/nova452/ComfyUI-Conditioning-Rebalance (Apache-2.0)
5. **comfyui-krea-vram-auto** (this listing's author) — the new VRAM-cap, auto-switch, and peak-stamp nodes — included as `comfyui-krea-vram-auto.zip` in this version's Download panel (MIT). Optional: `pip install comfy-aimdo` unlocks the hard VRAM-cap pin on the "Krea VRAM Cap" node — works fine without it too, just skips that one pin.

Install packs 1-3 via ComfyUI Manager search. Pack 4 isn't on the Manager registry — clone manually. Pack 5 ships with this listing — it's **not a link**, it's a separate file: on this page's right-side Download panel, open "Other Formats" and grab the **Archive** row (`comfyui-krea-vram-auto.zip`) alongside the Config/workflow JSON. Extract the folder into `ComfyUI/custom_nodes/`:
```
cd ComfyUI/custom_nodes
git clone https://github.com/nova452/ComfyUI-Conditioning-Rebalance.git
# then extract comfyui-krea-vram-auto.zip here (from this version's Download panel → Other Formats → Archive)
```

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🚀 **How to Use**

1. Install all model files + 5 custom node packs (above)
2. Load this workflow JSON into ComfyUI
3. Choose mode: **Mode Switch = 0** (keyword→auto-expand) or **1** (reference image→style capture)
4. Choose orientation: **Latent Switch = 0** (landscape) or **1** (portrait)
5. Leave VRAM Cap, Upscale Auto-Switch, and Quality-Tier Auto-Switch on **Auto** — they configure themselves from your GPU's total VRAM
6. Optional toggles: Detail Rebalance (default On), NSFW prompt style (default Off)
7. Queue → generate. Check console for the peak-VRAM proof stamp after each run.

Manual override available on both auto-switches if you want to force a branch regardless of detected VRAM — set the switch node's own `mode` widget to Force Index 0 / Force Index 1.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚙️ **Settings & Parameters**

• **Sampler** — euler / simple / CFG 1.0 / 10 steps (fixed, do not raise CFG above 2.0)
• **Mode Switch** — 0 (keyword) / 1 (reference image)
• **Latent Switch** — 0 (landscape) / 1 (portrait)
• **VRAM Cap target_usable_gb** — 6.5 (default; lower = more aggressive headroom, higher = allows more resident memory)
• **Upscale Auto-Switch threshold_gb** — 10.0 (total VRAM cutoff; raise this if you want upscale to only run on bigger cards)
• **Quality-Tier Auto-Switch threshold_gb** — 10.0, matched to the Upscale switch (raise/lower together to keep both decisions consistent)
• **Rebalance Multiplier** — 1.0 (default On)
• **NSFW Prompt Toggle** — 0 off / 1 on (default Off)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

💡 **Performance Tips**

• First queue is slow while Qwen3-VL-2B auto-downloads (~2.5GB)
• For predictable low-VRAM behavior across a long session, restart ComfyUI every so often rather than queuing many generations unattended — see Honest Disclosure above
• Keyword mode: start vague, let QwenVL expand it. Reference mode: sharp, well-composed references transfer style better

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

❓ **FAQ**

**"Where's comfyui-krea-vram-auto.zip? I don't see a link in the description."**
It's not a link — it's a separate file. On this page's right side, open the Download panel → "Other Formats" → grab the **Archive** row (`comfyui-krea-vram-auto.zip`), same place as the workflow JSON. Extract into `ComfyUI/custom_nodes/`, restart ComfyUI.

**"ComfyUI says I'm missing a node, Google has 0 results for it."**
Redownload `comfyui-krea-vram-auto.zip` from the Download panel above — an earlier upload was briefly missing one node class, fixed 2026-07-22. If you grabbed the zip before that date, just redownload and re-extract (same filename, new contents).

**"Do I need to install anything extra for the VRAM Cap node?"**
No, it runs out of the box. `pip install comfy-aimdo` is optional and only unlocks the hard VRAM ceiling — without it the node still runs a lighter unload/clear-cache pass automatically, no errors either way.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔗 **Also check out**

The earlier **v1 (SFW dual-mode)** and **v2 (NSFW + SeedVR2)** versions on this same model page, if you'd rather use a simpler single-purpose graph. Same dual-mode concept also available for **Z-Image-Turbo** (Apache-2.0) on my profile page.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📝 **Notes & AI Disclosure**

• **AI-Generated Content** — All example outputs are AI-generated by Krea-2 Turbo. Suitable for commercial and creative use (see licensing below)
• **Hardware Tested** — RTX 5080 16GB VRAM, CUDA 9.2+
• **VRAM Numbers** — all figures in this description are measured via `nvidia-smi`, not estimated. See Honest Disclosure section for exactly what was and wasn't tested.
• **Output Ownership** — You own all outputs. Commercial use OK if revenue < $1M/yr (Krea-2 Community License)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⭐ **Found this useful?**
• Like if it saved you time
• Comment your results — I read every one
• Follow for new ComfyUI workflows, all tested on 16 GB VRAM

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚖️ **Model & Node Attribution & Licensing**

**Krea-2 Turbo** (AlperKTS FP8 Quant)
• License: Krea-2 Community License — https://www.krea.ai/krea-2-licensing
• Commercial use: ✅ OK if total annual revenue < $1,000,000 USD

**Qwen3-VL-4B Text + Vision Encoder / Qwen Image VAE** (Alibaba Qwen)
• License: Apache 2.0 ✅ — Commercial use permitted

**SeedVR2 Upscaler** (ByteDance Seed Team)
• License: Apache 2.0 ✅ — Commercial use permitted
• Node: ComfyUI-SeedVR2_VideoUpscaler (numz) — Apache 2.0 ✅

**ComfyUI Custom Nodes**
• ComfyUI-QwenVL (1038lab): Apache 2.0 / BSD
• ComfyUI-Easy-Use (vjumpkung): Per upstream license
• ComfyUI-Conditioning-Rebalance (nova452): Apache 2.0 ✅
• comfyui-krea-vram-auto (Thinni63): MIT ✅ — new for this version

**Config credit**
• Rebalance node usage pattern adapted from DonutsDelivery's "Krea2 Turbo+NSFW patches+upscale v1.8" (Civitai), re-tuned/re-benched for this workflow

**Workflow JSON**
• License: CC0 Public Domain — free to use, modify, and redistribute without attribution (credit appreciated)

All example outputs are AI-generated. Model weights are third-party and covered by their respective licenses. Download weights separately from HuggingFace links above.