DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

Training a LoRA From Six Images in Seconds: What i2L Actually Produces

LoRA training usually means an hour of setup and a GPU running overnight. Image-to-LoRA collapses that: six reference images in, a working LoRA out, in seconds. The speed is real — the question is what you get for it.

The three graphs

Workflow Nodes Purpose
i2L training + test 23 Train and immediately render
i2L + ControlNet (combined) 62 Training with control guidance
i2L + ControlNet 43 Training with control guidance, lighter

The loader stack

Node Value
UNETLoader z_image_bf16.safetensors
CLIPLoader qwen_3_4b.safetensors, type lumina2
VAELoader ae.safetensors
LoraLoader [None, 1.1, 1.0]

An empty LoRA slot at weight 1.1

LoraLoader is loaded with None as the LoRA name and weight 1.1.

That is a placeholder: the graph is structured to receive the LoRA that i2L just trained, but no file is preloaded. The 1.1 weight is a telling detail — slight over-weighting on a freshly trained LoRA amplifies its effect, which is often what you want when the training data was small.

If you are running the training workflow as-is, the empty slot is correct. If you are using it as a template for a normal render, that 1.1 weight will apply to whatever you put there, and 1.1 on an art style LoRA is usually too strong.

The encoder type, again

qwen_3_4b.safetensors with type lumina2 here.

In the Z-Image ControlNet workflows in an adjacent release, the same encoder file is loaded with type qwen_image. Two different types, same file, same model family.

Same soft-failure pattern as elsewhere in this series: wrong type does not error, it just weakens prompt adherence. When merging nodes between Z-Image workflows, copy the loader rather than retyping it.

What i2L can and cannot learn from six images

Six images is enough to learn:

  • Colour palette and contrast range — the most transferable property
  • Line weight and edge treatment — especially for illustration
  • Overall composition tendency — centred versus off-centre, empty versus dense

Six images is not enough to learn:

  • A specific face — needs many angles of the same person
  • A precise object — needs views the six images probably do not have
  • Complex relational structure — "always a person holding an object on the left"

That split explains why i2L is best described as a style transfer tool. If you want to use it for identity, you will be disappointed, and not because the tool is bad.

The ControlNet combination graphs

62 and 43 nodes — meaningfully larger than the 23-node training graph, and the difference is the ControlNet path.

Why combine training with control: if the training images share a structure (all portraits, all centred), training against a control signal that encodes that structure produces a LoRA that generalises better to matching content.

EmptyLatentImage at 512×472 in both — small, because training does not need resolution.

Where this breaks

  1. Six images sets the ceiling. More data does not help beyond what the reference set supports.
  2. The empty LoRA slot at 1.1 is a template trap.
  3. Encoder type inconsistency across Z-Image workflows.
  4. "Seconds" is training time, not quality. A fast-trained LoRA on six images has limited range, and it will not tell you that.
  5. No evaluation in the graphs. The 23-node version renders a test image, but whether it matches the reference style is your judgement.

Practical approach

Train on six images that all share the property you want to transfer. Render a test immediately (that is what the 23-node graph does). If the result carries the palette but not the line style, you need more images of the line style specifically — or a different six.

Workflows

Three graphs, published on RunningHub with invite code rh-v1111:

  • Z-Image i2L fast LoRA training + test
  • Z-Image i2L training + ControlNet (combined)
  • Z-Image i2L training + ControlNet

Files

https://huggingface.co/AI-KSK/ksk-video-179-comfyui-workflows

Notes

  • Workflow files only. Model weights are not included; each keeps its own licence.
  • Tutorial exports, not a product.
  • The "seconds" figure is the tool's stated training time, not a measurement from these workflows.
  • Independent AI-KSK release. Not affiliated with Tongyi or RunningHub.

Optional API platform - 88API (affiliate link)

Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.

https://88api.ai/sign-up?aff=usx3

Disclosure: I may earn a commission from eligible purchases after you register through this referral link.

Top comments (0)