Zhenzhen's AI Radar · Flux.3 Klein · 2026-11-02 Evening Report
This issue features 4 new versions: Strongly Recommended 0 · Recommended 0 · Worth Watching 3 · Listed Only 1
Editor's Note This issue reviews 4 new Flux.3 Klein versions. There are no S/A entries, and no top recommendation is set. The three AIKSK Klein 4K upscale and reconstruction workflows share the same image pipeline (16 steps, CFG 1, long side 3840): the single-image version and the folder batch version using local GGUF for captions are rated B. The RunningHub ZIP batch version only changes input/output and relies more on the platform, so it is downgraded to C. None of the three have visible examples. The Klein version of the Mecha Detail LoRA has clear trigger words and is free, but lacks parameters, so it is rated B.
Worth Watching (3)
[B] Mecha Engrave · Flux.3 Klein mecha-v1
LoRA · Port · Author Xabsurd · Published 11-02 17:08 · Downloads 7
Flux.3 Klein version of the Mecha Detail Enhancement LoRA: The author claims it thickens armor layers, enhances the three-dimensional feel of decorations like beast heads and spikes, and boosts glowing energy parts, making mecha images appear more substantial.
Maintained at B: Clear purpose, explicit trigger words, and free, but lacks parameters. This section's version lacks author reviews.
Images
Reasons for Recommendation
- Port: The same author released five versions on the same model page on the same day: Krea 2, MOSS Image, Anima, Qwen 2.1, and Flux.3 Klein, all sharing the trigger word nijimecha. The Flux.3 Klein version has no separate description and is not mentioned in the author's version ranking (only stating Krea 2 is best, MOSS second, and Qwen 2.1 worst)
- Serves mecha settings, sci-fi posters, and mecha illustrations; the author claims it works for both clean mecha and moss-covered abandoned mecha
- Free, 2 visible examples; this version has 7 downloads (indicating popularity only)
How to Use
- Load adapter_model.safetensors, include the trigger word nijimecha in the prompt before describing the mecha body; strength, steps, and CFG are not provided by the author
Risks and Limitations
- The author has not evaluated the Flux.3 Klein version's effectiveness, and differences from other base model versions are unknown
- The weight filename is the generic adapter_model.safetensors; you must rename it after downloading
- The license only allows on-site rental and requires attribution; only 2 examples are provided; this entry has not been tested
Trigger Words: nijimecha
[B] FLUX 3 Klein 4K Upscale and Restore — Single Image · v1.0
Workflow · New Model · Author AIKSK · Published 11-02 13:18 · Downloads 12
AIKSK's Flux.3 Klein single-image 4K upscale and reconstruction workflow: First upscale with RealESRGAN, then have an online large model generate a description of the original image, and use Klein to reconstruct it with a long side of 3840, including original/result comparison.
Maintained at B: The most suitable of the three to start with, with visible parameters and tutorial links; however, it relies on online services and has no visible results.
Images
Reasons for Recommendation
- New workflow: The three Klein 4K workflows released by the same author on the same day share the same image pipeline (model upscale → scale by longest side → ReferenceMultiLatent to maintain resolution → sampling → decode). This is the single-image version, suitable for checking the description and result of one image before deciding on batch processing
- The author's card specifies the sampling array as 16 steps, CFG 1, euler/simple, with a long side target of 3840, but also frankly states that a fixed 3840×2160 is not guaranteed, nor is aspect ratio preservation or quality improvement
- Includes a Bilibili tutorial video link; free
How to Use
- Import flux3Klein4KUpscaleAnd_v10.json, prepare flux3_image_klein_sd_int8_convrot, qwen3vl_4b_int8_convrot, flux2-vae, and RealESRGAN_x4plus.pth; place the original image in LoadImage 13
- The description node RHLLMChatNode 16 depends on the RunningHub service (using gemini-3.1-flash-lite-preview). Confirm service configuration and image upload permissions first, then check the results in the comparison node
Risks and Limitations
- Depends on the RunningHub online LLM service; account, billing, and data terms must be confirmed by the user
- The author states no inference verification was performed, and there are no visible examples (1 hidden); the license does not allow derivatives; this entry has not been tested
- The card image is an AI-generated diagram, not the actual model output; the author has no publicly available standard-level examples
[B] FLUX 3 Klein 4K Local Folder Batch with GGUF Captions · v1.0
Workflow · New Model · Author AIKSK · Published 11-02 13:19 · Downloads 9
Local folder batch version of the same Klein 4K reconstruction process: Captions are generated using a local GGUF vision-language model, eliminating reliance on online services. It reads folders image by image, upscales and reconstructs, and saves in batch.
Maintained at B: The only version not relying on online services, with clear differences; however, the local caption model is heavy, and the batch process is unverified.
Images
Reasons for Recommendation
- New workflow: Differences from the single-image version are in two places—the input is changed to BatchImageLoader to read local folders, and the caption node is changed to llama_cpp loading the local Huihui-Qwen3.8-27B-abliterated Q4_K GGUF and corresponding mmproj, enabling offline operation
- Serves the need for unified local upscale and reconstruction of a batch of images; the author recommends testing progress, failure handling, naming, and output types with a small batch first, and ensuring the output directory is different from the input directory
- The image pipeline is the same as the single-image version: 16 steps, CFG 1, euler/simple, long side target 3840
How to Use
- Import flux3Klein4KLocalFolder_v10.json, prepare Klein, Qwen3-VL encoder, flux2-vae, RealESRGAN_x4plus, as well as the 27B GGUF model and mmproj for captions; replace the example input folder path and run a few images first
Risks and Limitations
- The caption model is a 27B GGUF, with VRAM/memory usage far higher than the online version. The author notes that compatibility with the processor string Qwen3.8 must be verified by the user
- The author states that automatic batching has never been fully verified, and there are no visible examples; the license does not allow derivatives; this entry has not been tested
- The card image is an AI-generated diagram, not the actual model output; the author has no publicly available standard-level examples
Listed Only (1)
- FLUX 3 Klein 4K Smart Upscale — RunningHub ZIP Batch · v1.0 — Workflow: RunningHub ZIP batch version of the same Klein 4K reconstruction process: Upload an image ZIP archive, upscale and reconstruct image by image, then package into a ZIP output, attempting to preserve original filenames.
Data: Civitai Public API (Base Model Flux.3 Klein), inclusion window 11-02 10:15 → 11-02 21:01 (Beijing Time). Recommendation reasons reviewed by editors; download counts indicate popularity only.