━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✨ **Wan2.2 I2V Funnel — Krea2 Turbo Text → Animated Video, One Click**

ComfyUI · Krea-2 Community License + Apache-2.0 · Fast DiT Turbo → Wan2.2-I2V-A14B

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Type a text prompt, get a 5-second animated NSFW video — no separate tools, no manual image export/import, no juggling two workflows. This graph fuses two of my other releases into one funnel: **Krea2 Turbo** generates the starting still internally, a VRAM-unload bridge node clears the image model out of memory, then **Wan2.2-I2V-A14B** (4-step Seko-V1 distill) animates it — all in a single ComfyUI graph, one Queue press. Built-in **QwenVL auto-prompt enhancement** transforms short subject descriptions into rich photographic prompts. If you've used my Krea2-Turbo Dual workflow, this is the natural next step: same still-image engine, now with motion.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✨ **Features**

✅ **One Graph, Zero Handoff** — SUBJECT prompt in, animated video out. No LoadImage node, no exporting stills between workflows — Krea2 generates the frame internally and hands it straight to Wan2.2.

✅ **VRAM-Safe Bridge** — `easy cleanGpuUsed` node fully unloads the Krea2 T2I stage from VRAM before Wan2.2's I2V stage loads, so two large models never have to co-reside.

✅ **4-Step Distilled Video** — Wan2.2-I2V-A14B split into high-noise (steps 0→2) + low-noise (steps 2→4) KSamplerAdvanced passes with Seko-V1 LoRA — 4 total sampling steps for fast turnaround.

✅ **Separate Prompt Channels** — SUBJECT prompt (CLIPTextEncode) drives the Krea2 still; a distinct MOTION prompt (umT5 encoder, wired through WanImageToVideo) drives the animation — describe *what* and *how it moves* independently.

✅ **GGUF Quantized Wan2.2** — both High-Noise and Low-Noise UNets run as Q4_K_S GGUF, trading a little precision for meaningfully lower VRAM than full-precision weights.

✅ **5-Second Smooth Video Output** — native 81-frame generation passes through RIFE 2x frame interpolation (→ ~32fps smooth motion), exported as H.264 MP4 via VHS_VideoCombine, ready to post or feed into further editing.

✅ **Funnel Companion to Krea2-Turbo Dual & Z-Image-Turbo** — same house pattern (cleanGpuUsed bridge, Euler/simple sampler, dual-mode heritage) extended into video.

✅ **Auto-Prompt Enhancement** — built-in QwenVL node auto-expands short subject descriptions into rich, detailed photographic prompts (~90 words) — just type what you want to see, it does the rest.

✅ **Motion-Smoothed 32fps** — RIFE VFI 2x interpolation applied to native 81-frame generation for ultra-smooth motion, doubling effective frame rate without changing video duration.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📦 **Required Models** (9 files, ≈30 GB total)

**Krea2 T2I stage:**

• krea2_turbo_fp8.safetensors (~7 GB) — Krea2 Turbo DiT 12.9B UNet, FP8 quant by AlperKTS

• qwen3vl_4b_fp8_scaled.safetensors (~2–3 GB) — Krea2 text + vision encoder (Qwen3-VL-4B, FP8)

• qwen_image_vae.safetensors (~1 GB) — Krea2 image VAE

**Wan2.2 I2V stage:**

• Wan2.2-I2V-A14B-HighNoise-Q4_K_S.gguf (~8–9 GB) — high-noise UNet (GGUF quant)

• Wan2.2-I2V-A14B-LowNoise-Q4_K_S.gguf (~8–9 GB) — low-noise UNet (GGUF quant)

• wan2.2_i2v_A14b_high_noise_lora_rank64_seko_v1.safetensors (~300–600 MB) — high-noise 4-step distill LoRA (Seko-V1)

• wan2.2_i2v_A14b_low_noise_lora_rank64_seko_v1.safetensors (~300–600 MB) — low-noise 4-step distill LoRA (Seko-V1)

• umt5_xxl_fp8_e4m3fn_scaled.safetensors (~6–7 GB) — Wan text encoder (umT5-XXL, FP8)

• wan_2.1_vae.safetensors (~250 MB) — Wan video VAE

**Prompt Enhancement (auto-downloaded on first use):**

• Qwen3-VL-2B-Instruct (~2.5 GB, auto-cached after first run) — VLM text-to-detailed-prompt enhancer; downloaded automatically by ComfyUI-QwenVL custom node on first workflow run, not manually placed

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⬇️ **Download Links**

📁 **ComfyUI/models/unet/Krea2/**

• krea2_turbo_fp8.safetensors — https://huggingface.co/AlperKTS/Krea2_FP8

📁 **ComfyUI/models/clip/** (or text_encoders/)

• qwen3vl_4b_fp8_scaled.safetensors — https://huggingface.co/AlperKTS/Krea2_FP8

📁 **ComfyUI/models/vae/**

• qwen_image_vae.safetensors — https://huggingface.co/AlperKTS/Krea2_FP8

• wan_2.1_vae.safetensors — search HuggingFace for exact filename (official Wan2.1 VAE, several mirrors)

📁 **ComfyUI/models/unet/Wan2.2/**

• Wan2.2-I2V-A14B-HighNoise-Q4_K_S.gguf — search HuggingFace/Civitai for exact filename (community GGUF quant of https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B)

• Wan2.2-I2V-A14B-LowNoise-Q4_K_S.gguf — same source as above

📁 **ComfyUI/models/loras/**

• wan2.2_i2v_A14b_high_noise_lora_rank64_seko_v1.safetensors — search HuggingFace for exact filename (Seko-V1 4-step distill LoRA, lightx2v-org)

• wan2.2_i2v_A14b_low_noise_lora_rank64_seko_v1.safetensors — same source as above

📁 **ComfyUI/models/text_encoders/** (or clip/)

• umt5_xxl_fp8_e4m3fn_scaled.safetensors — search HuggingFace for exact filename (standard Wan2.2 umT5-XXL FP8 text encoder)

⚠️ *Filenames above are exactly what the workflow JSON expects in its loader nodes — match them exactly, or update the loader node if your local copy is named differently.*

**Auto-Downloaded Models (no manual placement needed):**

• **Qwen3-VL-2B-Instruct** — automatically downloaded by the ComfyUI-QwenVL custom node on first workflow run (~2.5 GB, cached afterward); the prompt enhancer node will fetch it if missing.

• **rife49.pth** (RIFE VFI checkpoint) — automatically downloaded by the ComfyUI-Frame-Interpolation pack on first run; stored in the pack's ckpts/rife/ folder.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🧩 **Required Custom Nodes** (5 packs)

1. **ComfyUI-GGUF** — loads the Wan2.2 High-Noise/Low-Noise `.gguf` UNet files

2. **ComfyUI-VideoHelperSuite (VHS)** — provides the VHS_VideoCombine node (final MP4 export)

3. **ComfyUI-Easy-Use** (vjumpkung fork specifically) — provides the `easy cleanGpuUsed` VRAM-unload bridge node. Use ComfyUI Manager → "Install Missing Custom Nodes" to pull the correct fork automatically.

4. **ComfyUI-QwenVL** (1038lab / AILab) — provides the `AILab_QwenVL_PromptEnhancer` node for auto-expanding short prompts into rich descriptions. License: Apache-2.0/BSD.

5. **ComfyUI-Frame-Interpolation** (Fannovel16) — provides the RIFE VFI (video frame interpolation) node for smooth motion. License: MIT.

Install via ComfyUI Manager (search each pack name, or use "Install Missing Custom Nodes" after loading the JSON).

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🚀 **How to Use**

**Quick Start:**

1. Download all 9 model files → place in ComfyUI/models/ (unet/Krea2, unet/Wan2.2, clip, vae, loras, text_encoders — see paths above)

2. Install the 5 custom node packs via ComfyUI Manager

3. Load the workflow JSON into ComfyUI

4. Edit the **QwenVL Prompt Enhancer** node's `prompt_text` field — type a short subject description (e.g., "a cat in the rain"), and it auto-expands into a detailed photographic prompt wired to the SUBJECT CLIPTextEncode (do *not* edit the SUBJECT node directly — it gets overwritten each run)

5. Edit the **MOTION** prompt (CLIPTextEncode, umT5) — describes how the video should move

6. Queue → Krea2 renders the still → cleanGpuUsed unloads it → Wan2.2 animates it → RIFE interpolates to smooth 32fps → MP4 lands in your output folder

**What's happening under the hood:**

- Krea2 T2I: SUBJECT prompt (expanded by QwenVL) → EmptySD3LatentImage (960×640) → KSampler (euler, simple, 10 steps, cfg 1.0) → VAEDecode

- Bridge: `easy cleanGpuUsed` clears the Krea2 stage from VRAM

- Enhancement: QwenVL expands short SUBJECT prompt into ~90-word photographic description, wired to SUBJECT CLIPTextEncode

- Prep: ImageScale *up* to 1200×800 (from 960×640) for the video stage

- Wan2.2 I2V: WanImageToVideo (conditioned on the scaled still + MOTION prompt) → KSamplerAdvanced high-noise (steps 0→2) → KSamplerAdvanced low-noise (steps 2→4) → VAEDecode (Wan 2.1 VAE)

- Interpolation: RIFE VFI 2x frame interpolation (81 frames → ~161 frames)

- Output: VHS_VideoCombine → H.264 MP4, ~161 frames @ 32fps (5 seconds smooth motion)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚙️ **Settings & Parameters**

• **Krea2 Sampler** — euler, simple scheduler, 10 steps, CFG 1.0

• **Krea2 Resolution** — 960×640 (EmptySD3LatentImage), upscaled to 1200×800 before the video stage via ImageScale node

• **Wan2.2 Sampler** — dual KSamplerAdvanced: high-noise (steps 0→2) + low-noise (steps 2→4), 4 total steps, CFG 1.0

• **Distillation** — Seko-V1 4-step LoRA on both high-noise and low-noise UNets (lightx2v-org)

• **UNet Precision** — Q4_K_S GGUF for both Wan2.2 stages (lower VRAM vs. full precision)

• **Prompt Enhancement** — QwenVL (Qwen3-VL-2B-Instruct, 8-bit) auto-expands SUBJECT prompt into detailed photographic prompt (~90 words)

• **Frame Interpolation** — RIFE VFI 2x (doubles frame rate from ~16fps native to ~32fps smooth)

• **Output** — 81 native frames → RIFE 2x → ~161 frames @ 32fps → 5-second H.264 MP4 via VHS_VideoCombine

• **VRAM Bridge** — `easy cleanGpuUsed` between the Krea2 and Wan2.2 stages (do not remove — this is what keeps the two stages from co-residing in VRAM)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🎨 **Showcase**

Three sample clips from this funnel, each individually content-rated on Civitai — full SUBJECT + MOTION prompt pair for each is in its Generation Data. Two of three reach **full nudity**; one stays at a sheer-tease tier — the built-in QwenVL auto-prompt enhancement drives all three, no manual prompt-engineering needed, just type a short subject and queue:

• **Italian rooftop, golden hour** — sheer white dress, windswept side-profile, terracotta rooftops in soft-focus background. (sheer-tease tier, no nudity)

`SUBJECT: an Italian woman in a sheer dress, arching her back with wind-blown hair on a rooftop terrace at sunset`

• **Rain-streaked window, evening** — intimate over-the-shoulder glance, full nudity, soft indoor light.

`SUBJECT: a Japanese woman, fully nude, kneeling by a rain-streaked window in soft evening light, glancing back over her shoulder`

• **City balcony, golden hour** — quiet knee-hug pose, full nudity, skyline in soft-focus background.

`SUBJECT: an African woman, fully nude, sitting on a sunlit balcony ledge, city skyline in soft-focus background`

`MOTION (all three): wind gusts strongly, fabric whips and billows dramatically, hair blows wildly across face, dynamic motion, professional stock footage`

All three run the funnel at default settings end-to-end (960×640 Krea2 still → 1200×800 upscale → Wan2.2-I2V-A14B → RIFE 2x → 32fps MP4), one Queue press per clip — same reliability whether your subject lands sheer-tease or full-nude tier.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

💡 **Performance Tips**

• **Minimum Spec** — 16 GB VRAM **and** ~32 GB system RAM. `cleanGpuUsed` offloads the Krea2 stage to system RAM before Wan2.2 loads; low-RAM systems may thrash during the handoff.

• **VRAM** — Tested on RTX 5080 16GB — peak **15.9 GB** (15914 MiB), confirmed stable across repeated back-to-back runs with no restart. **Only ~470 MB headroom** — this is now a hard minimum spec (no comfort margin). Allocators running at ceiling.

• **Cold Start** — On first workflow run, both QwenVL (Qwen3-VL-2B-Instruct, ~2.5 GB) and RIFE (rife49.pth, ~50 MB) will auto-download and cache. Plan an extra 30–60 seconds the very first time; subsequent runs skip these downloads.

• **Expected Runtime** — roughly 3–4 minutes per clip on RTX 5080 — a Krea2 still generation plus a 4-step Wan2.2 animation pass plus RIFE interpolation (measured, not estimated).

• **Prompt Split** — keep SUBJECT (appearance/scene) and MOTION (camera/movement) prompts distinct; blending them into one prompt tends to confuse the Wan2.2 stage.

• **GGUF Quant Tradeoff** — Q4_K_S keeps VRAM down but is a lossy quant; if you have VRAM to spare, a higher-precision Wan2.2 quant will look sharper.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📝 **Notes & AI Disclosure**

• **AI-Generated Content** — All example outputs are AI-generated by Krea2 Turbo + Wan2.2-I2V-A14B. Respect local AI disclosure laws when publishing.

• **Hardware** — Tested on RTX 5080 16GB — peak VRAM 15.9GB (15914 MiB), verified across multiple consecutive runs with no restart (see Performance Tips above).

• **Configuration Only** — no model weights embedded in the JSON; download all 9 files separately from the sources listed above.

• **Workflow Reuse** — feel free to modify, share, and fork.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔗 **Also check out**

🔗 **Sister workflow:** [Krea2-Turbo Dual](https://civarchive.com/models/2736816) — the still-image engine this funnel is built on; QwenVL dual-mode prompt, same VRAM-safe cleanGpuUsed pattern.

🔗 **Sister workflow:** Z-Image-Turbo v2 (model 2742740) — a different still-image engine (Apache-2.0 S3-DiT) with SeedVR2 upscale toggle and triple orientation, if you want stills without the video stage.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⭐ **Found this useful?**

• Like if it saved you time animating stills

• Comment your results — I read every one

• Follow for new ComfyUI workflows, all tested on 16 GB VRAM

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚖️ **Model 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

**Wan2.2-I2V-A14B** (Wan-AI / Alibaba)

• License: Apache-2.0 — https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B

• GGUF quantization + Seko-V1 4-step distill LoRA (lightx2v-org) are permitted community derivatives

**Supporting Components** — all Apache-2.0

• Qwen3-VL-4B (Krea2 text/vision encoder) • Qwen Image VAE • umT5-XXL FP8 (Wan text encoder) • Wan 2.1 VAE • Seko-V1 4-step distill LoRA (lightx2v-org) • Qwen3-VL-2B-Instruct (prompt enhancer model)

**NSFW generation permitted under both licenses.**

**ComfyUI Custom Nodes**

• ComfyUI-GGUF • ComfyUI-VideoHelperSuite • ComfyUI-Easy-Use (vjumpkung fork) • ComfyUI-QwenVL (1038lab) • ComfyUI-Frame-Interpolation (Fannovel16, MIT)

**Workflow JSON** — original work, free to use, modify, and redistribute.

Full attribution detail (sources, license text, verification dates) is maintained separately in this listing's ATTRIBUTION.md. All example outputs are AI-generated. Model weights remain the property of their respective owners; download separately from the official sources above.