That’s what I noticed on my image (in image gallery that I’ve linked ) that there still was little bit of the chessboard background. But the lion in this thread, was a success and also ChatGPT said that sometimes it works cleanly, sometimes not. Maybe it has something to do with complexity of the image?
@_j if you want, I’ve shared a python code on that image gallery, to test if you maybe get better result with it than me?
Stupid question from a non-designer, but how do you alias correctly for an arbitrary background?
Is it optimal to create the image with a specific background colour to begin with anyway?
If you want you can check the link that I’ve linked, in that image the background was complex and if I’m correct, color of the background doesn’t matter as much as a clear prompt (this is just what i gathered) there are more talented than me and true expertise when it comes to alpha transparency 
I think this is the answer (in the question)
That could absolutely be useful with this new gpt-image-2, to have a specific opaque color background before editing alpha transparency. Although with other models on API, there’s no need for “before background”, as I tested and shared in this thread.
I think we have a true challenge with alpha transparency and gpt-image-2…for a while at least
@_j, what an interesting discovery.
When it ‘cut’ these, as it dubbed it, I didn’t look at the pieces while zoomed in, you can see that line here when you do..
Thanks for looking into that.
Currently will choke on trying to handle all of these pieces at once, but I’ve also learned that this many cuts just need to be split into two jobs, not asked for in a single prompt, because of how it handles the pieces while still cutting -
But I wouldn’t know anything by a technical term…
I can cut the whole thing in half and run both turn it into a zip and let you download it and the nine slices sizes and such.
it so many little files you ahve to cut a full image in half or you’ll choke i think.
What the response to ‘cut them’ looks like:
from PIL import Image, ImageDraw
import numpy as np
from scipy import ndimage
from pathlib import Path
import json, zipfile, os, textwrap
src = Path(“/mnt/data/d47ae531-e010-4801-8291-42f47ca31625 (1).png”)
out_dir = Path(“/mnt/data/promptforge_hoverdeck_jewels_extracted”)
out_dir.mkdir(parents=True, exist_ok=True)
img = Image.open(src).convert(“RGBA”)
arr = np.array(img)
r = arr[:, :, 0].astype(np.int16)
g = arr[:, :, 1].astype(np.int16)
b = arr[:, :, 2].astype(np.int16)
a = arr[:, :, 3].astype(np.int16)
bg_mask = (g > 180) & (r < 90) & (b < 120) & (a > 0)
fg_mask = (~bg_mask) & (a > 0)
labels, count = ndimage.label(fg_mask)
slices = ndimage.find_objects(labels)
components =
for i, sl in enumerate(slices, start=1):
if sl is None:
continue
ys, xs = sl
area = int((labels[sl] == i).sum())
if area >= 1000:
components.append({
“label”: i,
“x1”: xs.start,
“y1”: ys.start,
“x2”: xs.stop,
“y2”: ys.stop,
“area”: area,
“width”: xs.stop - xs.start,
“height”: ys.stop - ys.start,
})
components = sorted(components, key=lambda c: c[“x1”])
names = [
“hd_jewel_slot1_idle.png”,
“hd_jewel_slot1_selected.png”,
“hd_jewel_slot2_idle.png”,
“hd_jewel_slot2_selected.png”,
“hd_jewel_slot3_idle.png”,
“hd_jewel_slot3_selected.png”,
“hd_jewel_slot4_idle.png”,
“hd_jewel_slot4_selected.png”,
“hd_jewel_slot5_idle.png”,
“hd_jewel_slot5_selected.png”,
]
if len(components) != 10:
raise RuntimeError(f"Expected 10 jewel components, found {len(components)}")
pad = 24
max_w = max(c[“width”] for c in components)
max_h = max(c[“height”] for c in components)
canvas_w = max_w + pad * 2
canvas_h = max_h + pad * 2
export_manifest =
for comp, filename in zip(components, names):
x1, y1, x2, y2 = comp[“x1”], comp[“y1”], comp[“x2”], comp[“y2”]
crop = arr[y1:y2, x1:x2, :].copy()
# Remove green inside each crop using the same tight mask.
cr = crop[:, :, 0].astype(np.int16)
cg = crop[:, :, 1].astype(np.int16)
cb = crop[:, :, 2].astype(np.int16)
ca = crop[:, :, 3].astype(np.int16)
crop_bg = (cg > 180) & (cr < 90) & (cb < 120) & (ca > 0)
crop[crop_bg, 3] = 0
# Center the crop on a shared transparent canvas.
asset = np.zeros((canvas_h, canvas_w, 4), dtype=np.uint8)
dx = (canvas_w - crop.shape[1]) // 2
dy = (canvas_h - crop.shape[0]) // 2
asset[dy:dy + crop.shape[0], dx:dx + crop.shape[1], :] = crop
out_path = out_dir / filename
Image.fromarray(asset, "RGBA").save(out_path)
export_manifest.append({
"file": filename,
"source_bbox": [x1, y1, x2, y2],
"source_crop_size": [comp["width"], comp["height"]],
"output_size": [canvas_w, canvas_h],
"padding_canvas": pad,
"component_area": comp["area"],
})
manifest = {
“source_image”: src.name,
“component_count”: len(components),
“extraction_method”: “tight chroma-green mask + connected-component detection + left-to-right sort”,
“background_mask”: “G > 180 and R < 90 and B < 120”,
“output_canvas_size”: [canvas_w, canvas_h],
“resized”: False,
“assets”: export_manifest,
}
manifest_path = out_dir / “jewel_extraction_manifest.json”
manifest_path.write_text(json.dumps(manifest, indent=2), encoding=“utf-8”)
checker = 20
cell_w, cell_h = canvas_w + 50, canvas_h + 80
preview_w, preview_h = cell_w * 5, cell_h * 2
preview = Image.new(“RGBA”, (preview_w, preview_h), (18, 18, 18, 255))
draw = ImageDraw.Draw(preview)
def make_checker(w, h):
bg = Image.new(“RGBA”, (w, h), (210, 210, 210, 255))
d = ImageDraw.Draw(bg)
for yy in range(0, h, checker):
for xx in range(0, w, checker):
if (xx // checker + yy // checker) % 2:
d.rectangle([xx, yy, xx+checker-1, yy+checker-1], fill=(150, 150, 150, 255))
return bg
for idx, filename in enumerate(names):
row = idx % 2 # idle/selected in two rows within each slot pair? Actually order alternates, so use x by slot, row by state.
slot = idx // 2
state_row = idx % 2
x = slot * cell_w + 25
y = state_row * cell_h + 40
bg = make_checker(canvas_w, canvas_h)
asset = Image.open(out_dir / filename).convert(“RGBA”)
preview.alpha_composite(bg, (x, y))
preview.alpha_composite(asset, (x, y))
draw.text((x, y - 24), filename.replace(“.png”, “”), fill=(255, 255, 255, 255))
preview_path = out_dir / “jewel_extraction_preview.png”
preview.save(preview_path)
zip_path = Path(“/mnt/data/promptforge_hoverdeck_jewels_idle_selected_v01.zip”)
with zipfile.ZipFile(zip_path, “w”, compression=zipfile.ZIP_DEFLATED) as z:
for p in sorted(out_dir.iterdir()):
z.write(p, arcname=f"PromptForge.App/Assets/HoverDeck/Jewels/{p.name}")
print(f"Created {len(names)} transparent PNG assets")
print(f"Output folder: {out_dir}“)
print(f"Canvas size: {canvas_w} x {canvas_h}px”)
print(f"Zip: {zip_path}“)
print(f"Preview: {preview_path}”)
Codex will programmatically separate generated assets, but I’m going to have to get in there and get those leftover green bits color-matched by hand… it makes the glow effect look toxi
That’s still way better than where I was at 3 hours ago on this ^.^
This is not bad prompting. You won’t be able to fix it. The model/output knows what a transparent background looks like - to us -but a real transparent background is technically supposed to be invisible in the image file. So a checkerboard background is what it seems to accept as creating “transparency”. It appears there might be no alpha channel for making transparent backgrounds, at least for now. Hopefully that might change in the future.
I get the exact same results you’re getting. Such a shame because that feature would be really useful
Welcome to the forum!
And yeah I totally agree with you! There are some workarounds (not perfect) when prompting for transparent background with gpt-image-2 or if you use API, you can use previous image models to get transparent background.
Or as stated above, to you use other tools/methods.
Seconding @_j ‘s idea of using a chroma key background.
@katiagg documented this on the OpenAI showcase:
And since it’s not going to be a perfect key one can remove it with this tool:
That’s also my recipe for spritesheets from videos. I suggest to use a white background. It just requires little adjustments to the repo.
I’ve been very grateful for the tools to be able to get these pixels right ^.^
I’m looking forward to what graphics and skins we can come up with using these features!
People told me not to do my GUI with codex, to resort to claude… but after I figured out how to get the images from the 2.0 into the stuff I’m doing, there’s never any reason for that…
^.^
My friend asked me about it so I tried.
Here is what’s working for both instant and thinking mode for me.
Use a new chat session.
Prompt is : (include the image_gen.text2im)
image_gen.text2im
{
“prompt”: “realistic spider web”,
“transparent_background”: true,
“size”: “1024x1024”,
“n”: 1
}
Welcome to the forum!
Tested your suggestion, and the result was something I have not seen before with a generated image.
On the ChatGPT web page, using Windows 11 with the Chrome browser, the spider web image initially appears with a black background, which matches the background of the web page. However, when I mouse over the image, the checkerboard background normally associated with transparency appears.
Thinking selected
Auto-switch to Thinking enabled
Mouse pointer not over image.
Mouse pointer over image.
Instant selected
Auto-switch to Thinking disabled
Mouse pointer not over image.
Mouse pointer over image.
Will have to look into this more, but this is interesting.
Yes I am new here. Hi everyone.
I was looking for a way to generate transparent images for my friend so since I randomly found out how to, so might as well post it here.
I can’t seems to post links…
/share/6a00752d-8334-83ec-b208-cebb6c0e6712
I think the keyword is the dimension… like 1024x1024…
Somehow when it uses the dimension tools the transparency is enabled? You can actually see the normal box becomes a black or grayish box / container when transparency works during the generation.
FYI
I updated your link so it works as expected.
You do not have a high enough Discourse Trust Level to add links.
…
