How to Colourise B/W Photos and Postcards
I have always marveled at the hand-tinted colourisation of old B/W photos. As a postcard collector I have a preference for coloured postcards. Among the many things AI can do is to assist the postal historian with quick and easy B/W image colourisation.
There are several products that you can use for FREE, chief of which are ChatGPT and Grok. Both have strengths and weaknesses. ChatGPT is good for generating a quick precis and postcard colourisation while Grok does the same with a better researched precis with references for factual accuracy. Both programs will create images on request. Grok is more liberal while ChatGPT behaves more like a PC nanny. This post is about using ChatGPT to colourise your old Black and White postcards, something that is remarkably easy to do.
ChatGPT. Circa 1903. Postcard, Unused. (J. C. Juta & Co., Grahamstown.)
There are 'issues' with this colourisation but given the postcard's complexity the result is renarkable.
This was achieved with the first attempt / instruction.
I have subscribed to ChatGPT and am amazed at how clever and, more importantly, how easy it is to use. However, its results are often mixed. With image creation it's first attempt is often its best result you're going to get. The difficulty one experiences attempting to edit and improve an image is hugely frustrating. However, automatic postcards colourisation is smoother if you do not have high expectations.
A postcard colourised with ChatGPT.
Despite a very high resolution scan, we cannor read the largely illegible names on the far commercial buildings.
Basically, if you can't read determine a feature in an image yourself, you cannot expect AI to do so either .... yet.
It has been confused by sky bleeding into the white border on the right.
AI has added an image description with spelling error bottom right.
A renarkable job but nevertheless there are things it cannot do and or gets wrong.
A Bit about ChatGPT
ChatGPT is free but as in all commercial things its paying customers take precedence. I bought a subscription in order to test it and use it fully whenever I wanted. So, thinking you might want to know a little bit more about ChatGPT (if you do not already use it), I asked it to tell you about itself. I asked it to 'Give me a precis of ChatGPT's capabilities in 200 words'. It replied:
ChatGPT is an advanced AI language model (currently GPT-5) developed by OpenAI, designed to understand and generate human-like text based on prompts. It can converse naturally, answer questions, and provide detailed explanations across a wide range of topics, from everyday advice to technical problem-solving.
Its core strength lies in natural language understanding, enabling it to interpret context, tone, and nuance. It can summarize, rewrite, or expand text; translate between languages; generate creative writing; draft professional documents; and assist with brainstorming. Beyond text generation, ChatGPT can analyze data, produce structured outputs (like tables or charts), and reason step-by-step through problems. It adapts explanations to different skill levels, supports coding in many programming languages, and can help debug or optimize code.
It is also connected to real-time web search, allowing it to fetch up-to-date information when needed. With image capabilities, it can interpret, describe, and modify visuals, as well as generate new images from descriptions.
Its versatility makes it a tool for learning, research, productivity, and creativity. However, it’s not infallible—it can make mistakes and does not have personal experiences or consciousness. Its value comes from combining speed, breadth of knowledge, and adaptability into a single, interactive assistant.
Using AI to give you a precis of a subject is probably one of its best uses. However, you do need to check out the results. What it does, I think, is scan the internet for articles on the subject you want, then it writes it up for you. (Thousands of kids are using it to do their homework. The problem is AI creates the same homework!) More importantly, if its source data is wrong, so is its advice.
Colourising B/W Photos and Postcards
Regarding its image colourisation capabilities, first scan an image, then upload it into ChatGPT along with an instruction on what you want it to do. Unsurprisingly, I have found I get better results from higher definition scans. I suggest you scan your Black and White / Greyscale images in at 600 dpi in 24 bit colour and save it asw a JPG file.
If your B/W image is not 'Real Photo' postcard but instead is one that has been printed with a screen ie. it is made up of many little dots, use a Descreen function when you scan the image. Descreen is a simple check-box function in most scanning software. The postcard of 'Church Square, Grahamstown', top above, was screened. I used a Medium Descreen setting to automatically remove most of the dots. It is possible that ChatGPT will remove them during colourisation by itself. I am very happy with the result I got by removing them first.
I recommend that you clean the scanned image up in a program like Photoshop using a Despeckle tool, etc., before you upload it to ChatGPT for conversion. Any tears, flaws, stains and speckles etc., that can be seen in your scan will be reproduced if you do not get rid of them first.
Now, go to the ChatGPT website. Type this link into your browser. https://chatgpt.com/
ChatGPT is free but it may recommend to you that you pat for it. Once you get past all that to the workscreen, you will be confronted with something like the box below. As this exercise is supposed to be conversational, (you can instruct by voice command also), ChatGPT will ask you something like "What can I help you with?" or "What's on your mond today?".
Simply type in something brief like "Colourise the attached Black and White postcard of Grahamstown, Eastern Province, South Africa in 1900". It may be important that you let it know where and when the subject existed in time and space as it may reference a variety of other colour images, some of which may be the same subject which its will then draw on them for its 'imspiration'. If you click on the + sign you can upload your scanned and tidied-up JPG file into ChatGPT. When you are ready to go, click on this icon.
Get ready to be surprised! ChatGPT usually needs to think about things for a minute or so before it starts working on the project. It then takes about five minutes more-or-less to generate the image.
Franschhoek Huguenot Memorial Monument.
After opening ChatGPT - its free but I have found it quicker paying for it - a box comes up that says '+ Ask anything'. The plus sign is for uploading images. Upload your scanned image by clicking on + and typing your question into the same box. I asked it to: "Colorize the postcard below. It is from Fransch Hoek, Western Cape South Africa. The postcard shows the grey-white granite Huguenot Memorial Monument, a well-known landmark captured in numerous photos." Pressing Return at this point starts the process off.
This is the postcard that I uploaded as a 300 dpi JPG. Below is the result I got about five minutes later. I think it is great! The only issue is that it was returned to me as a 72 dpi JPG. Oh dear, I will probably have to read the manual to see how to get a hi-res file.
I think this is rather good. I have many other examples that I have colorised. If this is interesting email me and and I will post more.
Savoy Bioscope, Rosebank, Cape Town.
The image below is of the Savoy Bioscope, Rosebank, Cape Town, that I haunted as a teenager on Saturday mornings, swapping comics outside beforehand. Below is the original B/W image taken from the internet. I think the colour conversion is pretty perfect.
Circa 1950. Savoy Bioscope, Rosebank, Cape Town.
Note that I tidied up the white spots in the Douglas Fir before colourisation.
The Savoy has a special place in my memories. This colourisation does my memories proud.
Grand Hotel, Pretoria
1909. PRETORIA '8 FEB 09' to GB. Grand Hotel (Transvaal). (Hallis & Co., Port Elizabeth.)
I can show more colourised images if you are interested. Why not post some of yours in a Reply below.
Some considerations
Colourisation of B/W images is straightforward. As you can see from these colourised postcards, you can get some useful results if you are not too picky. The bigger problem with AI arises when you make changes and tweaks necessary to get the results you want. This is most likely to happen when asking ChatGPT to create an image, like the one below.
This image was automatically created by ChatGPT for my 'K is for ....' Zoom display.
It refused to show naked buttocks, so our subject has not pulled his trousers down for his 'Kuk in die Mielies'.
Curiously his trousers have detached below the knee - these are the PC Wonders of AI!
When you request changes to a ChatGPT image it is liable to run off and do its own thing, spoiling a half-decent first result. It will do things that you do not want, like put in headings in where none exist. It also has a habit of making spelling errors in headings. It has become hilariously infamous for putting more than five fingers on people it creates in images you request and having clocks that show just one time, its preferred ten to ten! Like any infant, AI will with patience, time and guidance hopefully mature into something useful.
Below is an old, somewhat faded photo, not a postcard. It is a snaphsot taken at the very start of the South African War (ABW to some). It appears to be an everyday camp scene .... only it is not.
The comment on its reverse "before the battle, 20th October" suggests that this is more than likely the Battle of Elandslaagte in Natal, fought on 21st October 1899. The only Scottish regiment at the Battle of Elandslaagte was the 2nd Gordon Highlanders. They attacked the Boer position on October 21st alongside the 1st Devonshire Regiment and the 1st Manchesters, as well as the Imperial Light Horse.
Lets see what we can do with this. Sadly I only have a 300 dpi JPG file available to me right now. I would prefer to do the conversion at 600 dpi as this will give ChatGPT more detail fo work with. As soon as I find the photo I will scan it at a higher resolution and colourise it again using the same prompts. If the results are notably different I will post them here and if not, I will add a comment to update this.
1899. Snapshot photo. "Before the battle, 20th October,"
In the above snapshot:
- Both soldiers have whitewash on their hands, the stooping man has it on his left boot.
- The erect, standing soldier, right, is wearing his sun helmet back-to-front.
- The erect, standing man's right hand is partly concealed by the bending man's sun helmet.
In the above colourised image:
- Both soldiers have lost the whitewash on their hands and a left boot.
- The erect, standing soldier, right, is wearing his sun helmet correctly.
- His right hand is fully visible, no longer partly concealed by the bending man's sun helmet.
When I asked ChatGPT to colourise this photo, I specifically asked "colourise this photo of soldiers from the 2nd Gordon Highlanders whitewashing stones in South Africa". I did so in the hope that ChatGPT would be clever enough to colour the men's kilts in the correct tartan of their regiment. It seems it has attempted to do this. Below is an example of Gordon Highlander regimental tartan.
Google's AI says "During the Second Boer War (1899-1902), Gordon Highlander uniforms were adapted for active service, with soldiers wearing a khaki apron over their tartan kilt and a greyback shirt with tails to hide the distinct kilt tartan from Boer marksmen. The kilt itself was the standard Gordon Highlander tartan, with red and black diced hose tops, but the sporran was replaced by a pocket on the apron. Ordinary socks, brogues, and khaki spats to protect the ankles completed the active-service dress". It would seem perhaps that this early SAW photo of the Gordon Highlanders was taken before their uniforms "were adapted for active service" in the face of long-range rifle fire.
Here is a less successful colourisation. This was automatically done within about five minutes or less.
This image is of thirteen Transvaal Scottish 2nd Btn volunteers in Pretoria during the 1914 Rebellion. Is is badly faded. I spent some time with Photoshop adjusting Contrast and Lighting in a poor attempt to enhance this 300 dpi image. If I were to scan the image again at a higher resolution, like 600 dpi, I think we would see a better colourised result. Its only use here is as an instructional piece.
My instructions to ChatGPT for this image were simply "Colorise this faded photo of Transvaal Scottish 2nd Btn volunteers in Pretoria in 1914 during the Rebellion. Their uniform is standard khaki. Canvas bandoliers, brown leather belts with brass buckles.. The earth is rich orange red. Cloudless sky". Below is the result it gave me. It is unusable without further difficult changes. A better solution than fiddling with it further would be to scan it at a higher resolution and give it slightly better instructions.
Images of the military are difficult to colourise with accuracy because so many regimental uniforms had slight differences despite being in the same Army, an example being the Scots and their tartans. An issue that concerns me is the colour of the British Army boot. I have been led to believe that the British Army wore black boots. A bombardier (never a reliable source) told me this during my National Service in the SADF. Apparently, according to him, "the SADF was unique in the British Army for wearing brown boots". Can anyone confirm or deny this? Knowing little details like this could add further accuracy to this colourised image.