VNCCS QWEN Encoder

When AHEKOT was asked what actually powers the new VNCCS 3.0 pipeline, he didn't hedge: "it's almost pure Qwen Image Edit 2511 with loras." This node is where that happens. It's VNCCS's own wrapper around the multi-image conditioning encoder Qwen-Image-Edit needs - instead of wiring up a generic text-encode node plus separate reference-latent nodes for each image, this one node takes your prompt, up to three reference images with individual weights, and hands back ready-to-sample positive and negative conditioning plus a starting latent.

Why Qwen-Image-Edit, and why multiple reference images

Qwen-Image-Edit became the default open instruction-editing model largely because its Apache 2.0 license let a real LoRA ecosystem grow around it - pose LoRAs, camera-angle LoRAs, clothing-transfer LoRAs, the exact kind of task-specific tooling VNCCS's pipeline leans on. Its 2509 revision added native multi-image editing - up to three inputs, combining person+person, person+product, or person+scene - which is precisely what this node's image1/image2/image3 slots expose: feeding a character reference, a clothing reference, and a pose reference into one edit pass rather than chaining three separate generations.

The inputs that matter

Required: clip, prompt (your instruction, plain language rather than tag soup - "the positive prompt describing the desired output," per the node's own tooltip), and vae.

The rest is optional but does most of the actual work:

• image1/image2/image3 with matching image1_name/image2_name/image3_name (default "Picture 1", "Picture 2", "Picture 3") - the reference images and the labels used to refer to them inside the prompt construction, and weight1/weight2/weight3 (default 1, range 0–2) controlling how strongly each reference's latent influences the result.
• latent_image_index (1–3, default 1) - which of your reference images becomes the base latent generation actually starts from, rather than just an influence.
• target_size (default 1024) and crop_method (pad, center, or disabled) - the tooltip is direct about the tradeoff: with cropping disabled, aspect ratio is preserved and only total megapixels change; with it enabled, mismatched aspect ratios get cropped or padded to fit.
• qwen_2511 (default true) - the tooltip calls this "strongly recommended," enabling handling specific to the 2511 model revision. Leave it on unless you know you're running an older Qwen-Image-Edit checkpoint.
• background_color (default White) - used only to flatten transparent pixels before encoding. Since most VNCCS sprites are RGBA after a chroma-key or RMBG2 pass, this matters more than it looks: feeding a transparent PNG straight into the VAE without flattening it first produces garbage, so this field exists specifically to paper over that.

Outputs: positive and negative conditioning, plus a starting latent - everything a KSampler needs downstream, in one shot.

How to install it

Bundled with the main pack, but the model weights are the real download here:

• ComfyUI Manager - search VNCCS - Visual Novel Character Creation Suite, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/AHEKOT/ComfyUI_VNCCS.git, then cd ComfyUI_VNCCS_Utils && pip install -r requirements.txt, restart, and click "Install missing custom nodes" in Comfy Manager once.
• The Qwen-Image-Edit checkpoint itself comes through VNCCS's Control Center downloader, not a manual HuggingFace grab - pick a quantization tier (Q4 lightest, Q5 balanced, Q8 best quality) based on what your GPU can hold and hit the Download button.

Common issues

Output looks nothing like your reference images. Check weight1/weight2/weight3 first - a reference at weight 0 is effectively not being used no matter how carefully you described it in the prompt. Also check latent_image_index: if it's pointing at the wrong image, generation starts from a base you didn't intend.

Transparent PNG references produce muddy or discolored results. That's the flattening step - background_color fills transparent pixels before the VAE ever sees them, so a background color close to your character's own palette can bleed into the edges. Set it to something that contrasts, not necessarily white by default, if your reference has a lot of near-white detail.

Results drift from the character across a chain of edits. This isn't a VNCCS-specific bug - it's Qwen-Image-Edit's known structural limitation: the model re-emits the whole frame each pass rather than patching just the changed region, so small drift compounds the more edits you chain. The 2511 revision specifically targets this ("mitigate image drift" is one of its headline claims), which is why qwen_2511 should stay enabled, but it doesn't eliminate the effect entirely - expect to occasionally regenerate from an earlier, cleaner reference rather than editing an already-edited-several-times sprite.