DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at bilibili.com

FLUX 3 Klein Reconstruction: Single Images, ZIP Batches, and Local Folders

#comfyui #imageprocessing #workflows #automation

A single image can look convincing while a batch workflow still pairs the wrong filename with its result, repeats an input, or stops after the first item. Scaling an image workflow therefore requires more than replacing its upload node with a folder loader.

This collection supplies three FLUX 3 Klein reconstruction graphs: a single-image route using an RH LLM node, a RunningHub ZIP route, and a local-folder route with GGUF visual captioning. They contain 16, 17, and 19 nodes respectively. Their reconstruction paths are similar; their input handling, caption services, and output contracts are not.

This article inspects the exported nodes, links, and source notes. No new inference, remote API request, batch run, or quality measurement was performed. The transcript describes the creator's experience with image enhancement; it does not establish new test results for every exported version.

Understand the shared reconstruction path

Each graph references four common assets:

  • flux3_image_klein_sd_int8_convrot.safetensors
  • qwen3vl_4b_int8_convrot.safetensors, alongside the stored encoder-type string flux3
  • flux2-vae.safetensors
  • RealESRGAN_x4plus.pth

These are strings stored in loaders, not bundled weights or verified upstream provenance. Confirm the matching node implementations and compatible assets before queueing. Filenames do not certify training mechanisms, quantization behavior, hardware support, or redistribution rights.

The image passes through ImageUpscaleWithModel at node 15 and ImageScaleToMaxDimension at node 11, then enters ReferenceMultiLatent at node 4. That node also receives positive conditioning, a zeroed-conditioning counterpart, and the VAE. Its outputs supply the reconstruction sampler; the decoded result reaches the appropriate saver.

The scaling node stores ["lanczos", 3840]. Its title and note describe a 3840 long-edge target, while ReferenceMultiLatent stores keep resolution. Those saved descriptions do not establish actual dimensions, aspect-ratio behavior, sharpness, or a fixed 3840×2160 frame.

Sampler 7 stores [9527, "fixed", 16, 1, "euler", "simple", 1]. The titles and notes describe a 16-step/CFG1/Euler/simple preset, but these exports do not provide a complete named widget schema. Treat the note as an attributed configuration description and inspect the matching UI before interpreting every array position as a confirmed parameter.

Workflow 01 — Inspect one image before scaling up

Purpose and requirements. The single-image graph is the simplest route for reviewing the relationship between an input image, its generated caption, and the reconstruction. It needs the common assets, compatible nodes, and a configured RH LLM service.

Node 13 is LoadImage. Its selected example filename does not prove that an input image is included. It sends the source image to three destinations: the upscaler, the RH LLM's image1 input at node 16, and the comparison node's original-image side.

The captioning node therefore sees the source image, not the enlarged image after node 11. Its response enters the text input of CLIPTextEncode at node 8. The saved blank text at that encoder is not evidence of an empty effective prompt, because the text input is connected.

How to use it. Import the graph, resolve missing dependencies, and select an authorized image. Check the LLM service setup and whether the image may be sent to that service. The node stores google/gemini-3.1-flash-lite-preview; that identifier alone does not identify the actual endpoint, credential type, or price.

Review the returned caption before relying on its instructions. The prompts request preservation of composition and identity while allowing plausible detail enhancement. Those requests are not guarantees: newly generated detail may still be wrong. After a real run, compare the original input with decoded output at node 12 and inspect text, faces, repeated patterns, and actual dimensions. Save node 5 uses prefix AIKSK_FLUX3_KLEIN_FINAL_4K_SINGLE.

Limitation. A connected save node is not proof that a file was generated, and a prompt asking for fidelity is not a measured fidelity result.

Workflow 02 — Test the ZIP mapping, not just the image

Purpose and requirements. The ZIP graph replaces the single-image input with HAIGC_LoadImagesFromZip at node 17 and adds HAIGC_SaveImagesToZip at node 18. It still contains the RH LLM captioning path and common reconstruction assets.

The ZIP loader's image output feeds the upscaler, caption node, and original side of the comparer. The decoded result feeds the ordinary image saver and the ZIP saver. A separate connection passes the loader's naming-information output to the ZIP saver's naming input.

That connection expresses an intended association. It does not prove that every input is captioned separately or that names and images remain paired after list mapping. The source note explicitly says that behavior depends on the RunningHub instance's ZIP plugin and other node versions.

How to use it. Resolve the required HAIGC nodes and service configuration. Begin with a small, authorized test archive rather than an entire collection. The note recommends three images with mixed orientations. Give those images distinguishable names and contents, then check output count, image-to-name pairing, orientation, actual size, and failure handling.

Inspect both output routes. Node 5 uses prefix AIKSK_FLUX3_KLEIN_FINAL_RH_BATCH; node 18 is the ZIP output route. A successful individual image save does not, by itself, establish that the archive is complete or that its metadata is correct.

Limitation. Do not advertise automatic per-image processing, original-name preservation, or batch completion as verified until the corresponding checks have actually been performed. Running the graph may transmit multiple images and incur service charges; confirm those permissions and terms first.

Workflow 03 — Local-folder input with a GGUF caption path

Purpose and requirements. The third graph uses BatchImageLoader, llama_cpp_model_loader, llama_cpp_instruct_adv, llama_cpp_parameters, and BatchImageSaver. It does not contain the RH LLM node used in the other two graphs.

The local model loader stores Huihui-Qwen3.8-27B-abliterated-Q4_K.gguf, Qwen3.8-27B-mmproj-BF16.gguf, and a Qwen3.8 processor string. The source note explicitly asks the operator to confirm model/projector compatibility and processor support. Names alone do not establish that this pair will load successfully in the installed plugin.

The note identifies dchuk/comfyui-batch-image-processing, lihaoyun6/ComfyUI-llama-cpp_vlm, and rgthree-comfy as dependencies. These are source-provided project identifiers, not freshly verified installation instructions. No tested version lockfile or measured memory requirement accompanies the package.

Follow the actual branches. Loader 13 sends an image to the upscaler, original comparison input, and visual-caption node 19. The caption response connects to encoder 8. The encoder's stored example text therefore is not the whole effective prompt description: it has an incoming text connection.

The upscaler remains in the connected image path. Although the note calls preprocessing optional, no separate on/off switch is visible in this export. Do not describe the current graph as already bypassing that stage.

How to use it. Replace the example input-directory setting with a real, authorized folder. Confirm the model, projector, processor, and node versions before processing a collection. Inspect the output directory and collision behavior in the installed UI, and keep outputs outside the input folder so a later batch does not ingest its own results.

Use a small test set first. The note describes automatic queue progression, but this export does not prove that the queue advances, recovers from failures, or terminates correctly. Check a successful item, a failed item, and completion behavior before leaving a larger job running.

Check filename and type connections. Loader 13 supplies both output_base_name and output_file_type to saver 5 through links. The presence of a stored png string is therefore not a basis for promising that all outputs will be PNG. The note describes a suffix and rename policy; verify their effective behavior with the plugin version rather than inferring a full widget schema from the raw array.

A progress formatter receives index and total-count outputs. It is auxiliary status wiring, not proof that files were generated or saved correctly.

A useful acceptance checklist

For all three routes, record the actual input, caption, effective settings, installed versions, generated output, dimensions, and observed errors. Compare source and result for changed text, identity drift, altered composition, or invented detail. For batches, additionally verify item counts, pairing, filenames, file types, output location, and stopping behavior.

The local caption path is structurally different from the remote paths, but a static graph is not a network audit or an offline certification. Model acquisition and plugin behavior still require inspection in the intended environment.

No VRAM figures, speed ratios, pixel-perfect guarantees, or successful batch-run claims are supplied here. The “4K” names describe the project's target, not a measured acceptance result. Process only images you have permission to use; generating a more detailed version does not remove the rights associated with the original.

The supplied YouTube title describes a different five-in-one scope, while these files contain three reconstruction routes. That link is withheld rather than presented as a verified walkthrough of this exact package. Confirm the matching tutorial and asset rights before publication.

Workflow pages

Getting the workflows

Published by this channel. The graphs are free; weights, LoRAs and upscaling dependencies are separate and carry their own terms.

Top comments (0)