Bilibili tutorial video: https://www.bilibili.com/video/BV11xJo6nEFh/
# FLUX 3 Klein — Local-Folder Reconstruction with GGUF Captions
## What this workflow does
Inspect a local-folder caption-and-reconstruction graph without assuming its automatic queue or model compatibility is already verified.
## Who should use it
ComfyUI users able to resolve compatible node versions, provide authorized images, and inspect actual outputs. This is one independent workflow, not a verified result package or a new model checkpoint.
## Requirements
Common loader references: `flux3_image_klein_sd_int8_convrot.safetensors`, `qwen3vl_4b_int8_convrot.safetensors`, `flux2-vae.safetensors`, and `RealESRGAN_x4plus.pth`. Weights, example inputs, and a tested installation lockfile are not included.
The local loader stores Huihui-Qwen3.8-27B-abliterated-Q4_K.gguf, Qwen3.8-27B-mmproj-BF16.gguf, and processor string Qwen3.8; the note explicitly leaves compatibility to verification. The upscaler remains connected despite being described as optional. No offline certification or successful automatic batch run is claimed.
## Configuration and steps
BatchImageLoader 13 feeds the upscaler, comparer, and llama_cpp_instruct_adv 19. The local caption connects to encoder 8. Loader outputs also connect to BatchImageSaver 5 output_base_name and output_file_type; the stored png string does not establish the effective output type.
Confirm the GGUF model, projector, processor support, and node versions. Replace the example input folder, inspect the output directory and collision behavior, and keep outputs outside the input folder. Test a small set for progression, failed-item handling, correct naming/type, and termination. BatchProgressFormatter is auxiliary status wiring, not completion evidence.
The image path passes through model upscaling, ImageScaleToMaxDimension, ReferenceMultiLatent, sampling, and VAE decoding. The scaling node stores ["lanczos", 3840], and its title/note describe a long-edge target. ReferenceMultiLatent stores keep resolution. The sampler array is [9527, "fixed", 16, 1, "euler", "simple", 1]; the source titles/notes call it a 16-step/CFG1 preset, but the export lacks a complete named widget schema. Confirm the matching UI instead of inventing parameter mappings. No fixed 3840×2160 size, unchanged aspect ratio, improved quality, or successful output is verified.
## Files
- [Workflow JSON]()
- [Promotional cover]()
## Version note and limitations
The JSON retains the supplied source bytes. This copy does not announce a new publication, model version, performance upgrade, or relationship to an existing remote project. Cover artwork is not a verified inference sample; supply a real reviewed sample separately if required.
No new runtime, GPU, cost, quality, or batch-completion evidence is included. A requested caption, stored target, or progress value is not an output guarantee. Verify installation, image-transfer permissions, effective naming/type settings, and actual results before deployment.
## Credits, tutorial, and rights
The workflow is supplied through AI-KSK with source notes retained. A matching international tutorial URL has not been confirmed from the supplied metadata; no replacement is guessed. Verify rights for workflows, node code, weights, cover artwork, and source images separately. Regeneration does not remove source-image obligations or create blanket redistribution permission.
Description
Bilibili tutorial video: https://www.bilibili.com/video/BV11xJo6nEFh/
# FLUX 3 Klein — Local-Folder Reconstruction with GGUF Captions
## What this workflow does
Inspect a local-folder caption-and-reconstruction graph without assuming its automatic queue or model compatibility is already verified.
## Who should use it
ComfyUI users able to resolve compatible node versions, provide authorized images, and inspect actual outputs. This is one independent workflow, not a verified result package or a new model checkpoint.
## Requirements
Common loader references: `flux3_image_klein_sd_int8_convrot.safetensors`, `qwen3vl_4b_int8_convrot.safetensors`, `flux2-vae.safetensors`, and `RealESRGAN_x4plus.pth`. Weights, example inputs, and a tested installation lockfile are not included.
The local loader stores Huihui-Qwen3.8-27B-abliterated-Q4_K.gguf, Qwen3.8-27B-mmproj-BF16.gguf, and processor string Qwen3.8; the note explicitly leaves compatibility to verification. The upscaler remains connected despite being described as optional. No offline certification or successful automatic batch run is claimed.
## Configuration and steps
BatchImageLoader 13 feeds the upscaler, comparer, and llama_cpp_instruct_adv 19. The local caption connects to encoder 8. Loader outputs also connect to BatchImageSaver 5 output_base_name and output_file_type; the stored png string does not establish the effective output type.
Confirm the GGUF model, projector, processor support, and node versions. Replace the example input folder, inspect the output directory and collision behavior, and keep outputs outside the input folder. Test a small set for progression, failed-item handling, correct naming/type, and termination. BatchProgressFormatter is auxiliary status wiring, not completion evidence.
The image path passes through model upscaling, ImageScaleToMaxDimension, ReferenceMultiLatent, sampling, and VAE decoding. The scaling node stores ["lanczos", 3840], and its title/note describe a long-edge target. ReferenceMultiLatent stores keep resolution. The sampler array is [9527, "fixed", 16, 1, "euler", "simple", 1]; the source titles/notes call it a 16-step/CFG1 preset, but the export lacks a complete named widget schema. Confirm the matching UI instead of inventing parameter mappings. No fixed 3840×2160 size, unchanged aspect ratio, improved quality, or successful output is verified.
## Files
- [Workflow JSON]()
- [Promotional cover]()
## Version note and limitations
The JSON retains the supplied source bytes. This copy does not announce a new publication, model version, performance upgrade, or relationship to an existing remote project. Cover artwork is not a verified inference sample; supply a real reviewed sample separately if required.
No new runtime, GPU, cost, quality, or batch-completion evidence is included. A requested caption, stored target, or progress value is not an output guarantee. Verify installation, image-transfer permissions, effective naming/type settings, and actual results before deployment.
## Credits, tutorial, and rights
The workflow is supplied through AI-KSK with source notes retained. A matching international tutorial URL has not been confirmed from the supplied metadata; no replacement is guessed. Verify rights for workflows, node code, weights, cover artwork, and source images separately. Regeneration does not remove source-image obligations or create blanket redistribution permission.