[HN Gopher] Pico-Banana-400k
___________________________________________________________________
Pico-Banana-400k
Author : dvrp
Score : 346 points
Date : 2025-10-26 02:01 UTC (20 hours ago)
(HTM) web link (github.com)
(TXT) w3m dump (github.com)
| vunderba wrote:
| From the paper
|
| _> The pipeline (bottom) shows how diverse OpenImages inputs are
| edited using Nano-Banana and quality-filtered by Gemini-2.5-Pro,
| with failed attempts automatically retried._
|
| Pretty interesting. I run a fairly comprehensive image-comparison
| site for SOTA generative AI in text-to-image and editing.
| Managing it manually got pretty tiring, so a while back I put
| together a small program that takes a given starting prompt, a
| list of GenAI models, and a max number of retries which does
| something similar.
|
| It generates and evaluates images using a separate multimodal AI,
| and then rewrites failed prompts automatically repeating up to a
| set limit.
|
| It's not perfect (nine pointed star example in particular) - but
| often times the "recognition aspect of a multimodal model" is
| superior to its generative capabilities so you can run it in a
| sort of _REPL_ until you get the desired outcome.
|
| https://genai-showdown.specr.net/image-editing
| typpilol wrote:
| I love your site I stumble across it once a month it seems.
|
| Or there's another very similar site. But I'm pretty sure it's
| yours
| vunderba wrote:
| Thanks! It's probably the same site. It used to only be a
| showdown of text-to-image models ( _Flux, Imagen, Midjourney,
| etc_ ), but once there was a decent number of image-to-image
| models ( _Kontext, Seedream, Nano-Banana_ ) I added a nav bar
| at the top so I could do similar comparisons for image
| editing.
| typpilol wrote:
| Yes that was exactly it.
|
| How often do you update it? It seems like something new
| every time I check. Or I forget everything..
| vunderba wrote:
| Honestly it's kind of inconsistent. Model releases
| sometimes seem to come in flurries - (it felt like
| Seedream and Nano-banana were within a few weeks of each
| other for example) and then the site will receive a
| pretty big update.
| lukasb wrote:
| What do you use for evaluation? gemini-2.5-pro is at the top of
| MMLU and has been best for me but always looking for better.
| vunderba wrote:
| Recently I've found myself getting the evaluation
| simultaneously from to OpenAI gpt-5, Gemini 2.5 Pro, and
| Qwen3 VL to give it a kind of "voting system". Purely
| anecdotal but I do find that Gemini is the most consistent of
| the three.
| lukasb wrote:
| Interesting, I'll give voting a shot, thanks.
| motbus3 wrote:
| I am running similar experiment but so far, changing the
| seed of openai seems to give similar results. Which if that
| confirms, is concerning to me on how sensitive it could be
| dangoodmanUT wrote:
| I found the opposite. GPT-5 is better at judging along a
| true gradient of scores, while Gemini loves to pick 100%,
| 20%, 10%, 5%, or 0%. Like you never get a 87% score.
| svantana wrote:
| That's a great website! Feature request: a button to toggle all
| the sliders left or right at the same time - would make it
| easier to glance the results without lots of finicky mouse
| moves.
| MattRix wrote:
| Seconding this. Once you've seen the original image once, you
| don't need to see it each time. The idea of syncing the
| sliders in the current group is a clever solution.
| vunderba wrote:
| Thanks. That's a great idea - I also incorporated @MattRix
| proposal of syncing the sliders. It should be up now!
| scotty79 wrote:
| Seedream seems to be clear winner
| vednig wrote:
| Other Post: https://news.ycombinator.com/item?id=45708493
| cjrd wrote:
| Eh
| gus_massa wrote:
| It looks like a post about the presentation in the
| conference. No discussion. Sometimes the first post about a
| topic doesn't geht traction but a layer post gets more
| popular.
| daemonologist wrote:
| I confess that I don't quite get the point here - is it just that
| they've paid the inference costs for a dataset than can be used
| for distillation/other research?
| peddling-brink wrote:
| Essentially yes, it's a data set that can help train or fine
| tune another model or similar research. From the site:
|
| > Pico-Banana-400K serves as a versatile resource for advancing
| controllable and instruction-aware image editing. Beyond
| single-step editing, the dataset enables multi-turn,
| conversational editing and reward-based training paradigms.
| TechSquidTV wrote:
| Can it be? Has Apple FINALLY joined the party? Very ironic they
| are using an open dataset from Google... and Gemini for prompts
| by Google.
|
| I'm happy to see something from Apple but this seems so low-tech
| that it could be one of my own local ComfyUI workflows.
| echelon wrote:
| They're distilling Nano Banana with a Google dataset, letting
| anyone more easily build and test their own systems. It's kind
| of funny how easy this is to do.
|
| "You wouldn't steal a car," but anyone can distill an
| expensive, fully trained model in order to build their own.
|
| This is going to be one of the most important categories of
| image model. It's good that we have more than Google and the
| Chinese (ByteDance, et al) with competent editing models. I
| don't think Flux Kontext is keeping up.
|
| It'd be really nice if we had a Nano Banana-calibur model as
| open source.
| kranke155 wrote:
| Flux Kontext is not keeping up. Even then Flux has become
| only partially open source. They keep the more advanced
| models API only.
| ttul wrote:
| Flux backbone is too rigid. Very difficult to fine tune.
| Qwen is where it's at these days.
| ThrowawayTestr wrote:
| I've gotten mind blowing results with flux.1 Dev. Is the
| API even better?
| TechSquidTV wrote:
| Qwen Image Edit? Though it is a little soft and plasticy
| ttul wrote:
| Unless you fine tune it... the guys of Qwen are amazing.
| skissane wrote:
| The license is CC BY-NC-ND - I'm not sure who is going to be able
| to use it given the NC-ND part... especially given the potential
| uncertainty over what uses count as commercial and what counts as
| derivative works. OTOH, given the bulk of this dataset is AI
| outputs, its copyrightability is an open question.
| hsbauauvhabzb wrote:
| I find caring about a licence for an LLM highly ironic.
| littlestymaar wrote:
| It's not even an LLM, it's a _dataset_.
|
| And clearly, if training on copyrighted material is fair use
| as every LLM makers claim, then this license has literally no
| weight.
|
| Also, NAL but IIRC an automatically generated dataset isn't
| copyrightable in the first place.
| DrewADesign wrote:
| Honor among thieves?
| niek_pas wrote:
| > CC-BY-NC-ND or Creative Commons Attribution NonCommercial
| NoDerivs, is the most restrictive license offered by Creative
| Commons. With this license, the user (while attributing the
| original creator) can only share the work but not change it in
| any way or ever use it commercially.
| qoez wrote:
| Output from a generative AI model has already been deemed non
| copyrightable and the license can't really overwrite that
| poly2it wrote:
| Worldwide?
| skissane wrote:
| I don't think anyone really knows the answer yet. UK law
| has much looser standards for copyrightability than US law
| - UK law accepts the "sweat of the brow" doctrine - mere
| human effort is enough to create copyright, even if it
| lacks any significant creative element-under UK law, a
| transcriptionist transcribing an audio recording creates a
| new copyright in the transcription separate from the
| copyright in the audio itself; US law does not consider a
| mere verbatim transcription to be sufficiently original to
| create a new copyright. But, will UK judges extend "sweat
| of the brow" to include AI sweat as well as human sweat? My
| gut feel is probably "yes", but I'm not aware of any case
| law on the topic yet. A complicating factor is there are a
| lot of wealthy vested interests who are going to be pushing
| for the law in this area to evolve in a way which suits
| them - both in the courts and in Parliament - so the law
| might not evolve in the way you'd expect if judges were
| just left to logically extend existing precedents.
|
| Even in the US, I think the situation is complex. If I
| prompt an LLM to edit a copyrighted human-written text, the
| LLM output is going to be copyrighted, because even if the
| LLM's changes aren't copyrightable, the underlying text is.
| And what happens if an LLM proposes edits, and then a human
| uses their own judgement to decide which LLM edits to
| accept and which not to? That act of human judgement might
| provide grounds for copyrightability which weren't present
| in the raw LLM output.
| BarakWidawsky wrote:
| Looks like the dataset is distilled from Gemini nano-banana
|
| Definitely very useful, but I'm so curious how the original
| datasets from these image editing models were created. I'm
| guessing a lot of it is synthetic data to construct scenes
| programmatically with layers
| ttul wrote:
| My rough guess is that they set a few workflows combining
| analytical and ML-based image manipulations to generate the
| training set. For instance, you can get a long way by having a
| segmentation model identify and mask various objects and then
| apply simple analytical manipulations to the masked areas such
| as changing their color, or diffusing new content into that
| area using masked guidance to another image diffusion model. In
| this way, you can create training pairs that your editing model
| learns to invert, such as "turn the woman's hair into blonde
| hair" (start with a blonde haired woman, mask the hair, and get
| a diffusion model to turn it brown; this gives you the scene
| you can now invert as a training pair).
| zuInnp wrote:
| Maybe it is only me, but all the emojis in the readme look like
| AI wrote it and instantly make stop reading it ...
| thinkingemote wrote:
| Another glaring giveaway is the over use of numbered lists and
| bullet point lists.
|
| Personally it makes me less likely to read it but the content
| might be useful. I have some general tech interest but am not
| overwhelmingly interested in the subject. Sometimes good things
| crop up on HN too.
|
| Now, if an author was writing for an audience with the
| intention to attract the interest of people who were not
| enthusiasts to become enthusiasts of their product they would
| create something readable and attractive. The LLM hasn't here.
|
| Together, this leads me to think that the readme is not for me
| but is just for dedicated enthusiasts.
| stefan_ wrote:
| All the READMEs these days are such a tell. It's okay when
| explicitly prompted, but now thanks to reinforcement learning
| through people who have no clue, all the models just top off
| every change with some pointless documentation change.
| ThrowawayTestr wrote:
| I've seen so many repos with bare readmes that I don't even
| mind generated ones.
| cubefox wrote:
| Any idea why they didn't use GPT-4o image generation?
| Alifatisk wrote:
| I think it's because Geminis nano banano is better than 4o
| imagegen at creating and editing images from instructions
| Jackson__ wrote:
| Because GPT-4o is too orange (literally).
| neom wrote:
| Someone who works in AI told me they think that was trained
| in as a "watermark", apparently the same is true with the em-
| dashes, to "ease people into AI" or something.
| svantana wrote:
| Gemini is better. See results of double-blind evaluation here:
|
| https://lmarena.ai/leaderboard/image-edit
| pwython wrote:
| Valid question, as they already have a partnership with OpenAI
| to use ChatGPT in Siri. I personally use GPT for illustrations
| and Nano Banana for photo edits (Midjourney for realistic
| photos).
|
| As an aside, perhaps they're using GPT/Codex for coding. Did
| anyone else notice the use of emojis and - in their code?
| whywhywhywhy wrote:
| 4o generates new images that try to be close but not exact it
| doesn't edit existing ones so wouldn't be much use for this.
| sollewitt wrote:
| AI industry: please _please_ get it together with naming. There
| shouldn't be this much overlap between this, a dataset, and a
| massive image model which was already given a garbage name to
| begin with.
|
| Don't get me started in how "agent" is a term of art that means
| absolutely nothing, encompassing everything from a plain old
| shell script to a full language model.
| 3abiton wrote:
| To be fair, as the "AI" industry came late, the dibs were
| called on all the cool acronyms/names (LoRa for example).
| andai wrote:
| That's true. And if they did somehow think of a cool name,
| they simply wouldn't have the resources to purchase the
| rights to it ;)
| NBJack wrote:
| Is it just me, or do they switch the names a bit as they go
| along? Maybe I just missed something?
|
| > Dataset Statistics
|
| > Nano-Banana-400K contains ~400K image editing data, covering
| a wide visual and semantic range drawn from real-world imagery.
| fritzo wrote:
| Missed opportunity to call it banana-seeds-400k
| MeteorMarc wrote:
| Thought this was about Raspberry Pi because of the assocations
| with the banana pi and the pi pico.
| 3form wrote:
| And 400. I genuinely thought it's going to be a Banana Pi in
| keyboard form factor.
| ttul wrote:
| Image editing model training is fascinating. One method for
| training image editing models involves using a second model to
| apply the inverse of the change you want the model to learn.
| Typically, the task you're asking the second model to perform is
| easy, whereas the inverse task is difficult.
|
| For example, you might ask the second model to cover the person's
| face with a black square; a VLM model notes that the person is a
| man with brown hair and round glasses. Then, during training, the
| resulting image is presented along with the prompt, "Remove the
| black square from the man's face. He has brown hair and round
| glasses."
|
| The model now learns how to remove black squares and replace them
| with a man's face with brown hair and round glasses.
|
| Since the training data is easily synthesized using existing
| models, you can generate enormous amounts of it - often very
| cheaply. For specialized editing tasks, this technique is really
| powerful. Build your training set for your special purpose task,
| fine tune an existing image editing model such as Qwen Image Edit
| to produce a new checkpoint or LoRA (often a LoRA is more than
| good enough) and then you have a special purpose model to perform
| whatever narrow editing task you need it to perform on your image
| data.
| onlyrealcuzzo wrote:
| Are these models built atop models that already understand
| natural language?
|
| If the commands all follow the same syntax, it's easy to
| imagine how you can generate a good training set.
|
| But how to they fully grasp natural language to be able to
| perform tasks worded unexpectedly, which would be easy to
| parse, if they understood natural language?
| jerf wrote:
| "But how to they fully grasp natural language to be able to
| perform tasks worded unexpectedly, which would be easy to
| parse, if they understood natural language?"
|
| A Large Language Model. Pardon me for spelling out the full
| acronym, but it is what it is for a reason.
|
| I think a lot of the whiz-bang applications of LLMs have
| drowned it out, but LLMs are effectively the solution to the
| long-standing problem of natural language understanding, and
| that alone would be enough to make them a ground-breaking
| technology. Taking English text and translating it with very
| high fidelity into the vector space these models understand
| is amazing and I think somewhat underappreciated.
| ThrowawayTestr wrote:
| I love open datasets. The future of LLM is open source models.
___________________________________________________________________
(page generated 2025-10-26 23:00 UTC)