跳到正文
DEV Community· Jerry Simth·· 2 小時前AI 評分57

Ideogram 4.6 評測:文字呈現、原生 4K 與兩輪編輯實測

I tested Ideogram 4.6: exact text, real 4K, and a second edit that changed only the coat

AI 導讀

作者以六次生成、共 70 credits、約 $0.70 實測 Ideogram 4.6,測試文字呈現、原生 4K、兩輪編輯及透明 PNG,並稱六次均通過。4K 生成檔為 4096×4096;第二輪編輯只把外套改成深紅色,保留人臉、姿勢、海灘背景及 1760×2336 尺寸。

正文

I run a small studio site that hosts 37 AI image models side by side, and I test each one the same way: real prompts, first output kept, no re-rolls. Ideogram 4.6 launched November 19 with three headline claims — the best text rendering in the business, native 4K output, and precise multi-step editing — so I ran six generations against them on November 22. All six passed. The most interesting result was not the text, it was the second edit.

The test ran November 22, 2026: six generations, 70 credits, about $0.70 in total. Ideogram 4.6 costs 10 credits a picture at 1K and 2K, 20 at 4K. Edits are 10 credits at 1K and 2K, 33 at 4K, with up to four reference images included free. Here is what came back.

Text, twice: a poster and a label

"Movie poster for a fictional jazz festival, art deco style, a trumpet player silhouette against a deep blue night sky with stars, large bold title text at the top reading "BLUE NOTE NIGHTS", smaller text below reading "JAZZ FESTIVAL · JUNE 12-14", elegant gold lettering, vintage poster design"

Art deco jazz festival poster with the headline BLUE NOTE NIGHTS and the date line JAZZ FESTIVAL · JUNE 12-14, both spelled exactly, generated with Ideogram 4.6

"BLUE NOTE NIGHTS" in large gold type, "JAZZ FESTIVAL · JUNE 12-14" underneath — both spelled exactly, middle dot and en dash included, first output. One honest detail: the model stacked the title across two lines rather than one, so the layout is its choice, not the brief's. The words are right, which is the part that used to break.

Then the harder text test — small type on a curved surface:

"Coffee bag packaging design, matte black coffee bag with a minimal cream label, large text reading "SUMMIT ROAST", smaller text below reading "ETHIOPIA SINGLE ORIGIN", coffee beans scattered beside the bag, studio product photography, warm light"

"SUMMIT ROAST" as the brand line, "ETHIOPIA SINGLE ORIGIN" below it, both exact, first output. Two text tests, zero spelling errors. This is the capability Ideogram is known for, and 4.6 keeps it.

Native 4K, measured

"Photorealistic alpine lake at sunrise, mirror-still water reflecting snow-capped peaks, a wooden cabin with warm lights on the shore, pine forest, mist over the water, extremely detailed, crisp textures"

Alpine lake at sunrise with a lit wooden cabin on the shore, rendered at native 4K, generated with Ideogram 4.6

The output file measured 4096×4096 pixels. That matters because "4K" in image tools often means a 2K render passed through an upscaler; this one was generated at full size, so the cabin lights, the pine texture and the mist hold up at a size you can crop hard or print large. At 20 credits it costs double the 2K tier — worth it for finals, overkill for drafts.

The edit, twice: only the coat changed

Round 1 took a rooftop portrait and swapped the background for a sunset beach, keeping the face, pose and dark coat. Round 2 ran on round 1's output with one instruction: change the coat to deep red, keep everything else exactly the same.

"Change the person's dark coat to a deep red coat, keep the face, pose, sunset beach background and everything else exactly the same"

The same portrait after a second edit: the coat is now deep red, face, pose and sunset beach background unchanged, edited with Ideogram 4.6

Only the coat changed. Same face, same pose, same beach, same 1760×2336 dimensions — even the footprints in the sand are still there. Two rounds in, nothing has drifted. That is what "precise editing" has to mean in practice. Ideogram says 4.6 ranks first on Rapidata's multi-step image edit benchmark at 1,892; that is Ideogram's own claim, and the lead over 4.5 (1,866) sits within the margin of error, so I read it as "at least as good as 4.5" rather than a proven leap. Two rounds on one portrait are a check, not a benchmark — but it is the check that matters for real revision workflows.

Transparent PNG, measured

"A red vintage scooter in side view, studio product photograph, clean isolated object, sharp details, soft shadow beneath"

Red vintage scooter on a transparent background, generated with Ideogram 4.6

A real alpha-channel PNG, not a white background: I measured the alpha channel pixel by pixel and 82.9% of the frame is fully transparent, the rest is the scooter. No background-removal step, no fringing to clean up. (The full measurement method, including the red-background check, is in my transparent PNG guide.)

What I actually learned

  1. Text is table stakes now, and 4.6 clears it twice. Specified text on a poster and on packaging, both exact, first try. The test I have not run yet is invented text — a credits block the model writes itself — which broke MiniMax H3 and which Grok passed. That is the obvious next test for 4.6.
  2. The two-round edit is the real result. Single edits are easy to demo; second edits are where models drift. 4.6 changed one thing and kept everything else, which is what revision workflows actually need.
  3. 4K is real, and priced like it. Native 4096×4096 at 20 credits for generation is fair; the 4K edit at 33 credits is the price to notice before you click it. Draft at 2K, save 4K for the final.
  4. Know the limits. This test ran at Medium quality, the only tier on the studio — other quality settings were not tested. And the 4K edit tier is priced but I kept no 4K edit sample here, so that tier is untested in this post.

The full test — all six generations with verdicts, the prompts verbatim, and the pricing — is on the blog post, and the studio itself is at pizyn.com if you want to run your own prompts (20 free credits at signup).

All images in this post are first, unedited outputs of the prompts above, generated November 22, 2026. Six generations, 70 credits, no re-rolls.

來源:DEV Community · dev.to