IAMCCS NextFrameBuilder

IAMCCS_NextFrameBuilder looks like a single node with a big pile of widgets. It isn't. At execution time it expands into an entire Qwen Image Edit 2511 sampling chain - model loader, three LoRA applications, conditioning, KSampler, tiled decode - running behind a storyboard-friendly face. That's the trick, and it's worth knowing before you judge it by its intimidating input list: the node is a workflow wrapper, the author's words, "a professional next-scene storyboard UI backed by a dynamic Qwen graph."

What does that get you? A storyboard loop that actually works. You give it a source image (the current scene), describe the next scene, and it generates a frame that continues the scene - preserving identity, wardrobe, lighting, location geometry - then hands you a committed image you can feed straight back in as the source for the scene after. Two previews, next-scene generation, commit-and-reuse. No re-wiring a dozen nodes between every shot.

Mechanically it builds the chain with ComfyUI's graph-builder API rather than hard-coding node calls, which is why it can offer both a GGUF and a native-UNET backend from one dropdown. The chain it assembles is exactly the Qwen Image Edit reference stack: either UnetLoaderGGUF (default model qwen-image-edit-2511-Q4_0.gguf) or UNETLoader with the bf16 safetensors; then up to three LoRAs applied model-only - the Lightning 4-step LoRA (default strength 1.0), a next-scene_lora-v2-3000 continuation LoRA (0.8), and a Qwen 2511 light-and-shadow LoRA (0.8). From there it's ModelSamplingAuraFlow with the shift you set (3.1 default), a CFGNorm wrapper, TextEncodeQwenImageEditPlus for positive and negative with your reference images scaled to conditioning_megapixels, a FluxKontextMultiReferenceLatentMethod choosing your reference_method, and a KSampler running the defaults you see on the node - 4 steps at CFG 1.0, euler/simple, full denoise. Decode is tiled (VAEDecodeTiled, default 512px tiles) precisely because the default output is 1920×1088 and nobody wants that in one unfragmented decode. The output frame is committed through the companion IAMCCS_NextFrameCommitPreview machinery, which is what turns it into a reusable file.

The inputs that actually matter

• source_image - a file that must already exist in ComfyUI's input folder. The node validates this up front and refuses to run with a clear message otherwise.
• prompt_text - pre-filled with a strong continuity prompt ("Next Scene: The camera moves slightly forward to a medium shot… Preserve the exact identity, face, hairstyle, wardrobe, body proportions, location geometry…"). Keep its structure; that's what holds character across shots.
• run_token - the manual "force rerun" switch. Bump it to regenerate.
• model_loader + the model filename widgets - GGUF or native backend, and which files to use. The defaults point at specific filenames; your first task is making sure those exact files exist in your models folders (diffusion_models, clip, vae, loras).
• width/height, seed/seed_mode, steps, cfg, sampler_name/scheduler/denoise - the sampling knobs, with defaults tuned to the Lightning LoRA's 4-step fast path.
• Optional reference_image_2 / reference_image_3 for multi-frame conditioning.

Outputs: next_frame (IMAGE), storyboard_json (the running board), generated_filename (the committed file - feed this back as the next source_image), and cine_linx (for the bridge into video).

The honest take

This is the highest-convenience node in the pack, and also the most demanding on your model collection: Qwen Image Edit 2511 is a 20B-class model, so you'll want the Q4 GGUF or a fat GPU, plus the Qwen CLIP and VAE and three specific LoRAs. The reference workflow names exact files like next-scene_lora-v2-3000.safetensors that aren't in the repo - you source those yourself, and mismatched names fail loudly at validation, not silently. Where people get burned: forgetting source_image must already live in the input folder (a text path, not an image node output - this wrapper manages files, not tensors), and cranking steps/quality knobs without realizing the defaults are tuned for the 4-step Lightning path.

Install

The pack once:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes

or via ComfyUI Manager (search "IAMCCS"), then restart. Then chase the models: Qwen Image Edit 2511 (GGUF or native), the Qwen 2.5 VL CLIP, the Qwen image VAE, and the Lightning / next-scene / light-and-shadow LoRAs, plus ComfyUI-GGUF if you use the GGUF loader. The README's floor of a current ComfyUI, Python ≥ 3.12 and a recent PyTorch applies.